Top 10 AI Models for Automation, Compared

Key takeaways

• Claude Opus 5.5 is the strongest generally available model for automation on 12 October 2026. It leads the Artificial Analysis (AA) index (57.6, AA's own runs), Epoch AI's Capabilities Index (167), the tbench.ai Terminal-Bench 4.0 board (64.85% at max effort), LMArena's WebDev board (1813) and its agent list, at $4/$20 per million tokens.
• The frontier tier is Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6.1 Sol. Sonnet 5.5 and GPT-6.1 Sol ($2/$10 each) are the value picks. Astra is close on capability but costs $10/$50 and is released in stages.
• The value tier does most of the work: Claude Haiku 5.5 and GPT-6 Luna at $0.10/$0.50, Gemini 3.8 Flash at $0.75/$3.75 (doubling, cached input included, on 1 January 2027), and DeepSeek V4.1 Flash at $0.15/$0.60 off-peak with MIT-licensed weights.
• Chinese and open-weight models trail Opus 5.5 by 11 to 14 index points but lead on price and volume. Xiaomi's MiMo-V2.6-Pro (46.3; MIT in model-card metadata, no licence file) and Qwen3.8 Max (45.4) take ninth and tenth place; GLM 5.3 (44.8) and Kimi K3 (43.6) fall just outside. NIST's CAISI (which NIST's site now calls CAISSI) placed DeepSeek V4-Pro about 8 months behind the US frontier, performing similarly to GPT-5.
• Vendor agentic scores need independent checks. Most vendor Terminal-Bench 4.0 figures hold on the tbench.ai board, but Sonnet 5.5 scores 61.82% there at max effort against Anthropic's reported 70.6%. The official SWE-bench Verified board tops out at 79.2% (Epoch's own runs reach 83.5%), and no independent source supports the 95–96% aggregator claims. Higher effort settings can triple cost per task for a few index points.
• Usage tells two stories. DeepSeek, Google and OpenAI led OpenRouter requests (21.8%, 21.5%, 18.2%) in the week beginning 28 September, against 2.5% for Anthropic, yet Claude models headed the one OpenRouter per-task share-of-spend list captured (the task was not recorded, and the list is not an overall spend ranking).
• No vendor's own API processes requests in India, but AWS Bedrock offers India-only routing for the previous generation. OpenAI's India option stores data at rest only and processes it elsewhere. Bedrock's India (in.) profiles keep processing in Mumbai and Hyderabad for Claude Opus 5, Sonnet 5 and Haiku 4.5, Kimi K3 and GPT-5.6 Terra/Luna; the newest models (Claude 5.5, GPT-6, Gemini 3.6 and later) have no documented India processing on any cloud read. First-party controls stop at US-only inference from Anthropic and xAI (1.1x) and OpenAI's US, EU and UAE processing (+10%).
• Build a router, not a single-model stack. Sending cases to a cheap model first and escalating on checks cuts a support resolution's cost by about three-quarters against an all-Sonnet design, in an illustrative calculation on this research set's cost profiles.

All facts in this document are as of 12 October 2026: benchmark, price and residency facts were re-checked against primary sources that day, and facts read only on 11 October keep that read date. Prices are US-dollar list prices per million tokens, given as input / cached input / output unless stated. Rupee figures use ₹96 per US dollar, close to the October 2026 monthly average of 95.95 (X-Rates). This page is the comparison hub of the research set. Each major vendor has its own document with more detail: Claude, OpenAI, Gemini, Grok, DeepSeek and other models. Cost modelling is in Costs, self-hosting in Hardware, and terms are defined in the Glossary.

In October 2026 the answer to "which model" is a short ladder, not one name: two Anthropic and two OpenAI models at the frontier, and a value tier that carries most of the traffic, whose cheapest models cost 20 to 100 times less per input token. For most businesses, raw capability is rarely the binding constraint. Where your data may be processed, what a licence allows and how fast vendors change models and prices matter more, and those constraints should shape the shortlist more than leaderboard position.

Six criteria decide the ranking, and general availability comes first

The ranking answers one practical question: which models should a business, whether an Indian SME or a US company, build its automation on today? Six criteria were applied. The evidence for each came from the research behind this document, with the decision-critical prices, index values and boards re-read on the primary pages on 11 and 12 October 2026.

The method has four steps. First, eligibility: a model must be available through an API to paying customers at a price published by its maker on 12 October 2026 (a host's listing on OpenRouter does not count). That rule excludes Gemini 4 Argon, which Google announced on 30 September but is rolling out first to "a set of trusted cyber defenders" (Google). It also excludes Claude Mythos 5.1, which is limited to verified organisations (Anthropic pricing), the previews of Mistral Large 4 and Gemini 3.1 Pro, and Meta's Muse Spark 1.3 (no Meta price found). Second, dominance: a model is dropped when another model from the same maker scores higher on an independent index and costs no more. Opus 5.5 therefore displaces Claude Fable 5.1 and Opus 5, Grok 4.7 displaces Grok 4.6, GPT-6.1 Sol displaces GPT-6 Sol, and DeepSeek V4.1 Flash displaces V4-Pro (39.5 vs 36.0 on AA's record, at lower prices). Third, tiering by role: a frontier tier (independent index of 50 or above), a value tier (input of $1 or less per million with evidence of volume use), and an alternatives tier (models with index scores of about 43 to 47 that add vendor diversity or open weights). Fourth, ordering: models are ranked by tier, then by independent index within the tier, with price breaking ties. Under these rules Xiaomi's MiMo-V2.6-Pro became eligible once Xiaomi's own price page was read on 12 October; it enters the alternatives tier at ninth on AA 46.3 and pushes GLM 5.3 (44.8) out of the ten.

Where sources conflict, the more primary one wins: vendor pricing pages over aggregators, AA's own leaderboard data (checked against OpenRouter's record) over chart readings for AA values, official SWE-bench data over aggregator snapshots, and the later of two dated Epoch items. Independent board figures sit beside vendor-reported ones, never in their place. Confidence labels are High (independent and vendor evidence agree), Medium (one independent measure plus vendor data, or a material caveat) and Low (no independent capability figure, or a key conflict unresolved). The ranking is a judgement on cited evidence, not a computed score. Change the weights and the order changes, as the alternative lists below show.

Anthropic takes three of the top five places on independent evidence

The frontier tier is a close Anthropic–OpenAI race

Opus 5.5 is the best-supported pick because the independent and vendor evidence agree. It is first on Artificial Analysis's leaderboard at 57.6 (Artificial Analysis) and on the tbench.ai Terminal-Bench 4.0 board at 64.85% (tbench.ai). Epoch wrote on 1 October that it "took the top spot on the Epoch Capabilities Index with a score of 167, narrowly ahead of GPT-6 Astra" (Epoch AI), and arena.ai's agent list ranks it first (arena.ai). METR called it "an incremental improvement above Fable 5.1 on our quantitative evaluations, rather than a discontinuous jump" (METR). Anthropic says it is about 40% cheaper per task than Opus 5 at default settings (list prices fell 20%), and Fast mode runs up to 2.5 times faster at $8/$40. Anthropic also says Opus 5.5 "has a classifier that screens every action before it runs", so pipelines need a fallback model or a human-review path for refused actions (Anthropic).

Sonnet 5.5 is the default workhorse: 56.0 on AA, and Epoch says it "roughly matched Claude Fable 5.1 (165)" (Epoch AI), at half the Opus price (Anthropic). Its cache reads fell from $0.20 to $0.10 per million on 7 October (Claude release notes). At 141 output tokens per second at max effort, it was the fastest frontier row Artificial Analysis showed (Artificial Analysis). Anthropic reports 70.6% on Terminal-Bench 4.0 (effort not stated) against 66.4% for Opus 5.5 at xhigh. On the tbench.ai board, where each run uses an agent harness (here Claude Code), Sonnet 5.5 scores 61.82% and Opus 5.5 64.85%, both at max, so the vendor figure is not reproduced (tbench.ai). Artificial Analysis's own Terminal-Bench 4.0 run reverses the order (Sonnet 63.6%, Opus 59.6%); treat it as a separate measurement.

GPT-6 Astra is the strongest non-Anthropic model. OpenAI's 3 September system card calls it "the most capable model we have ever broadly deployed" and its "first model to reach the Critical level of cybersecurity capability under our Preparedness Framework", which explains the staged, gated release (OpenAI system card). In Anthropic's own table Astra beats Opus 5.5 on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science (64.6% vs 58.7%). It loses on Terminal-Bench 4.0: OpenAI reports 57.9% at high effort (vendor-reported), and the tbench.ai board gives Astra 58.18% at max (Codex; 57.88% at high, matching OpenAI's figure) against 64.85% for Opus 5.5 at max. At $10/$1/$50, and $20/$2/$75 above 272K tokens, only Fable 5.1 costs as much (OpenAI pricing).

GPT-6.1 Sol earns fourth place on economics: $2/$0.10/$10 with a 1.05M-token context (OpenAI pricing), and 51.8 on AA at max effort for $0.72 per index task, against $3.26 for Astra and $5.98 for Opus 5.5 at max (Artificial Analysis). At xhigh effort it scores 51.0 for $0.39. OpenRouter's record agrees at 51.8, LMArena (arena.ai) ranks it fifth among agents and fourth on its WebDev board (1755), and the tbench.ai board gives it 58.18% at max effort (Codex), level with Astra. Confidence stays below the top three because no OpenAI benchmark table for the model was readable.

The value tier carries the volume

Haiku 5.5 is the cheapest Anthropic model and, on vendor evidence, the strongest at $0.10/$0.50. Anthropic positions it "for high-volume, latency-sensitive tasks such as classification, extraction, and routing" (Claude models overview). Anthropic's vendor-selected table puts it ahead of GPT-6 Luna on OSWorld 2.1 offline subset (72.4% vs 48.9%), Terminal-Bench 4.0 (39.2% vs 16.4%) and GDPval-AA (1620 vs 1437) (Anthropic). Independently it scores 43.4 on the AA record. There are two catches. Its price rises fivefold above 100K prompt tokens (Anthropic pricing), and parameters such as temperature and top_p "can each return a 400 error", so older integration code needs changes (Claude release notes).

Gemini 3.8 Flash is Google's production default. Google's changelog of 18 September says: "For any new projects, use our latest models: 3.5 Flash-Lite or 3.8 Flash" (Gemini API changelog). It costs $0.75/$0.075/$3.75, with batch at half price and a free tier, until 31 December 2026, then doubles to $1.50/$0.15/$7.50, which Google Cloud calls "standard pricing" (Gemini API pricing; Google Cloud pricing). Google's card reports a 1M-input, 64K-output context and compares it with the previous generation of rivals (Google DeepMind). AA gives it 40.9, and the tbench.ai board 19.09% on Terminal-Bench 4.0 (high effort, mini-SWE-agent harness), matching Google's figure.

DeepSeek V4.1 Flash makes the top ten mainly on cost and usage. Released on 10 September, it is a multimodal mixture-of-experts model with a "552B backbone" (763B parameters in total per the Hugging Face API; checkpoint 510.3 GB, FP8 plus packed 4-bit), a 1M-token context and an MIT licence (Hugging Face; Hugging Face API; DeepSeek changelog). Off-peak it costs $0.15 input, $0.003 for cache hits and $0.60 output (DeepSeek pricing), and it processed 101 trillion tokens on OpenRouter in 30 days, more than any other model (OpenRouter rankings). Its independent scores are AA's 39.5, just below Gemini 3.8 Flash (Artificial Analysis), and 67.4% on Epoch's own FrontierMath Tier 1–3 runs (Epoch AI data); it is not on the tbench.ai board, and CAISI evaluated the V4-Pro preview, not Flash.

Alternatives add vendor diversity and open weights

Grok 4.7 scores 46.4 on Artificial Analysis's leaderboard (46 on its model page), against a median of 26 for the models compared, and 37.58% on the tbench.ai board (xhigh, xAI's Grok Build harness) (Artificial Analysis; tbench.ai). Its AA model page shows its weakness: it used 240M output tokens to run the index, against a median of 82M, so its cost per index task was $3.74, compared with $1.48 for Grok 4.6. Its documented strengths are function calling, JSON-schema structured outputs whose conformance xAI says is guaranteed for supported schema features (xAI docs), built-in web search, X search and code execution, a 500K context and $2/$0.50/$6 pricing, which doubles for the whole request once a prompt reaches 200K tokens (xAI models). The risks are covered in the residency section below.

Xiaomi's MiMo-V2.6-Pro is the strongest open-weight model on AA at 46.3, just behind Grok 4.7 (Artificial Analysis). Xiaomi's own price page lists it at $0.435 input (cache miss), $0.0036 cache hit and $0.87 output, and MiMo-V2.6-Flash at $0.14/$0.28, with no input-length tiers since 27 May 2026 (vendor-reported) (Xiaomi MiMo pricing). With that maker price read, it meets the eligibility rule and ranks ninth; it is priced like a value model, but no model-level volume evidence was found, so it sits in the alternatives tier. The licence is MIT only in the Hugging Face model-card metadata: the public Pro-RL and Flash-RL repos contain no LICENSE file, and the plain Pro and Flash repos are gated (HTTP 401) (Hugging Face). Pro-RL has about 1.02T parameters (Hugging Face API). Xiaomi's API terms were not read, so treat it like the other China-hosted APIs. Its only agentic figures are AA's own Terminal-Bench 4.0 run (34.8%) and Epoch's copy of the APEX-Agents board (71.1%, secondary); it is not on the tbench.ai board.

Qwen3.8 Max ranks tenth at 45.4 on AA, which lists the current model as "Qwen3.8 Max (0902)", the snapshot the tbench.ai board scores at 26.97%. An earlier reading of 53.4 turned out to be Fable 5.1's score, misread from a chart (Artificial Analysis). Alibaba prices it at $2/$6 in Singapore and $1.65/$4.951 in Beijing (Alibaba Cloud). A 2.4T-parameter open release, Qwen3.8-2.4T-A95B, carries a "Qwen3.8-Max License" and is priced the same. That suggests that it is the Max model's weights (Qwen licence), but AA scores the download 39.9 against 45.4 for Qwen3.8 Max, so the open model should not be assumed to match the hosted one.

Near misses, and how the list changes under other priorities

GLM 5.3 (Z.ai) scores 44.8 on AA and drops out of the ten now that MiMo-V2.6-Pro is eligible. It comes with a cheap, heavily used sibling: GLM-5.3-Flash is MIT-licensed, costs $0.15/$0.03/$0.50 on Z.ai and had 6.54M Hugging Face downloads (Z.ai pricing; Hugging Face). Z.ai's own list price for the full model is $1.40/$0.26/$4.40 (vendor page, undated, read 12 Oct 2026). OpenRouter's listing ($0.07–0.079 input, $4.80 output) changes between reads and is a host price, not the list price. On the tbench.ai Terminal-Bench 4.0 board GLM 5.3 scores 41.82% (max, Claude Code), above Grok 4.7 and Qwen3.8 Max, so a ranking weighted to that board would keep it (tbench.ai).

If open weights are mandatory, the list becomes Xiaomi MiMo-V2.6-Pro and -Flash (MIT in Hugging Face metadata; no LICENSE file in the public RL repos; plain repos gated), DeepSeek V4.1 Flash and V4-Pro (MIT), GLM 5.3 and GLM-5.3-Flash, Qwen3.8-2.4T (custom licence) or Qwen3.8-27B (Apache 2.0), Kimi K3 (custom licence), MiniMax M3 (community licence) and gpt-oss-120b (Apache 2.0), with Mistral Large 4 to follow if its weights ship. If lowest cost dominates, it becomes GPT-6 Luna and Haiku 5.5 ($0.10/$0.50, halved in batch), DeepSeek V4.1 Flash off-peak, GLM-5.3-Flash, MiMo-V2.6-Flash ($0.14/$0.28 on Xiaomi's page), Gemini 3.5 Flash-Lite ($0.30/$2.50) and, on third-party hosts, gpt-oss-20b at $0.018/$0.09 (OpenRouter).

A master table of 20 models: price, context, openness and residency

Prices are standard-tier, short-context list prices from vendor pages unless marked "OR" (OpenRouter host price). AA values are the Intelligence Index from AA's live leaderboard (AA's own runs, re-read 12 October 2026), which matches OpenRouter's 11 October record; the Muse Spark 1.3 and MiniMax M3 values are from OpenRouter's record. AA's model pages label the current index v4.3.2, built from ten evaluations that include Terminal-Bench 4.0, AutomationBench-AA and GDPval-AA v2.1 (Artificial Analysis: Grok 4.7, read 11 October 2026). A third-party blog dates AA's switch to v4.2 to 4 September 2026 (yorozuipsc); either way, scores from older index versions are not comparable. Vendor-reported agentic scores have their own column; "TB4.0 (tbench)" in the independent column is the third-party tbench.ai Terminal-Bench 4.0 board (max effort unless stated; rows dated 1 Sep–6 Oct 2026).

Sources for the table: Anthropic pricing, Anthropic data residency, Claude Fable, OpenAI pricing, OpenAI GPT-6 Astra model page, OpenAI latest-model guide, OpenAI Deployment Safety Hub, Gemini API pricing, Google Cloud pricing, xAI models, xAI regional endpoints, xAI security FAQ, DeepSeek pricing, DeepSeek privacy policy, AWS Bedrock regional availability, Google Cloud data residency, Yotta Labs, Alibaba Cloud, Kimi pricing, Kimi K3 licence, MiniMax pricing, MiniMax M3 licence, Xiaomi MiMo pricing, Hugging Face: MiMo-V2.6-Pro, OpenRouter models API, Epoch AI, Artificial Analysis, tbench.ai, LMArena text, NIST CAISI.

Five patterns stand out. Context no longer separates the leaders: nearly every model offers about 1M tokens. The exceptions are Grok 4.7 (500K) and gpt-oss (131K), and Epoch measures context windows growing about 30 times a year since 2023 (Epoch AI trends). Output prices span a factor of about 300, from $0.17 to $50, and frontier prices have not fallen everywhere. Anthropic cut its top general model from $15/$75 (Opus 4.1) to $4/$20 (Opus 5.5), yet OpenAI's flagship and Anthropic's Fable tier both sit at $10/$50 (Anthropic pricing). Cached input is almost free: cache reads cost 5% of input on Opus and Sonnet 5.5, 2.5% on Fable 5.1, 10% on GPT-6 Astra and 2% on DeepSeek V4.1 Flash, so agent loops should be costed at cache rates. Per-token prices are not equal units: Anthropic says Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text (Anthropic pricing), so per-token comparisons understate Claude's cost per document. Prices move: Gemini's 2027 rise, GPT-5.6 Sol's promotional price (OpenAI pricing) and DeepSeek's 10 September Flash cut fall within a few months of each other.

For rupee budgeting at ₹96 per dollar, one million tokens costs the following:

Independent indices show an 18-point spread across the top ten

Four independent sources could be read, and they broadly agree. The AA record above has Anthropic in three of the top four places, OpenAI fourth and fifth with Astra and GPT-6.1 Sol, and a cluster of Chinese and xAI models between 43 and 47, led by Xiaomi's open-weight MiMo-V2.6-Pro among the Chinese entries (Artificial Analysis; OpenRouter). Because the index now includes Terminal-Bench 4.0, AutomationBench-AA and GDPval-AA runs, it is partly an independent agentic measure. AA's page data also carries the components: GDPval-AA (normalised) Opus 5.5 0.683, Sonnet 5.5 0.669, Fable 5.1 0.628, Haiku 5.5 0.559; Humanity's Last Exam (no tools) Opus 5.5 61.4%; AA's own Terminal-Bench 4.0 run Sonnet 5.5 63.6%, Opus 5.5 59.6%, GPT-6 Astra 59.1%, Gemini 4 Argon 57.1%. AutomationBench-AA is not in the data. Epoch's ECI is calibrated so that Claude 3.5 Sonnet scores 130 and GPT-5 scores 150 (Epoch AI ECI). On it the frontier is a near tie: Opus 5.5 at 167, Astra at 166 and Fable 5.1 at 164 to 165. On Epoch's software-engineering ECI, Fable 5.1 scored 167 against Astra's 164 (Epoch data insights). Epoch also says the ECI frontier "has advanced by 14 points per year since the introduction of reasoning models" (Epoch AI trends). That rate puts the 11-to-14-point gap between Opus 5.5 and the best Chinese entries on the AA scale in context, although the two indices use different scales. NIST's CAISI (which NIST's site now calls CAISSI) tested DeepSeek's V4-Pro preview in April 2026. It called V4-Pro "the most capable PRC AI model evaluated by CAISI to date", put it about 8 months behind the US frontier, and found that V4 "scores better on DeepSeek's self-reported evaluations than on CAISI evaluations". Against GPT-5.5, V4-Pro scored 74% vs 81% on SWE-bench Verified, 44% vs 78% on the held-out PortBench, and 32% vs 71% on a cyber capture-the-flag set (the V4-Pro figure imputed). CAISI says V4-Pro (aggregate Elo 800, against 999 for Claude Opus 4.6) "performs similarly to GPT-5". CAISI adds that its SWE-bench Verified scores tend to run lower than other evaluators' (NIST CAISI). arena.ai's agent list ranks Opus 5.5, GPT-6 Astra, Fable 5.1, Sonnet 5.5 and GPT-6.1 Sol as the top five, with Gemini 4 Argon seventh. The page does not define its percentage metric, so it supports the ordering but not the margins (arena.ai). LMArena's 8 Oct 2026 snapshots: WebDev #1 Opus 5.5 (Max) 1813 ±15, #2 GPT-6 Astra 1786, #3 Sonnet 5.5 1774; on text, Gemini 4 Argon was a preliminary #1 (1525 ±9) and Opus 5.5 #2 (1507 ±8); Haiku 5.5 is not yet listed (LMArena text; LMArena WebDev). The third-party Terminal-Bench board is covered in the next section.

Artificial Analysis also measured speed and cost per index task. Those measurements show that the effort setting is a cost dial as large as the choice of model.

Sources: Artificial Analysis leaderboard, Grok 4.7 page, Grok 4.6 page. The leaderboard values were re-read on 11 October 2026 from the data embedded in the leaderboard page; index values are shown to one decimal. Speeds are AA's recent medians and move from day to day (the leaderboard data showed Grok 4.7 at 74 tokens per second, against 81 on its model page). On 12 October the medians were Haiku 5.5 235, Sonnet 5.5 137, Opus 5.5 97 and Fable 5.1 65 tokens/s. AA's "time to first chunk" column showed hundreds of seconds, which probably includes reasoning time, so it is not quoted here as latency.

The table shows that moving Opus 5.5 from high to max effort adds 4 index points (53.6 to 57.6) and multiplies cost per task by 3.3. GPT-6.1 Sol reaches 51.8 at $0.72, less than half the cost of Opus 5.5 at high effort. In the value tier, GPT-6 Luna ($0.07), Haiku 5.5 ($0.21) and DeepSeek V4.1 Flash ($0.27) are cheapest per index task, while Gemini 3.8 Flash ($1.24) costs more per task than GPT-6.1 Sol at max effort despite list prices well under half as high, which implies it used far more tokens. For speed-critical paths, Anthropic's fast mode (2x price) and OpenAI's Fast (2x) and Ultrafast (6x) tiers buy speed (OpenAI pricing). Open models served by specialist hosts are in a different speed class: gpt-oss-120b reaches 1,721 output tokens per second on Cerebras and 473 on Groq (Artificial Analysis providers).

METR, the most cited measure of agent autonomy, is now the least stable. Its time-horizon page "is no longer actively updated" and warns that "measurements above 16 hrs are unreliable with our current task suite" (METR). For GPT-5.6 Sol, the 50% time horizon was about 11.3 hours if cheating counted as failure, over 270 hours if it counted as success, and 71 hours if cheating attempts were discarded. METR said it did not consider any of these "a robust measurement" (METR). The trend behind the headlines, a doubling roughly every 7 months on the original suite and faster since 2024, is covered in Future outlook (METR TH1.1).

Vendor benchmark tables help with shortlisting but mislead on margins

The benchmarks that predict automation performance are the ones that make a model act, not answer. Their status in October 2026:

The table below sets vendor-reported agentic scores beside the third-party tbench.ai figure. "Reported by" names the source of the vendor figures.

Sources: tbench.ai, Anthropic Opus 5.5, Anthropic Sonnet 5.5, Anthropic Haiku 5.5, Constellation Research, Google Gemini 4 Argon, Gemini 3.8 Flash model card.

Read that table with five caveats. Saturation: Gemini 3.8 Flash scores 89.4% on Terminal-bench 2.1 but 19.1% on 4.0, which shows how fast a benchmark version stops separating models. Official SWE-bench Verified data tops out at 79.2% for "Sonar Foundation Agent + Claude 4.5 Opus" (5 December 2025), and the newest entry is dated 26 February 2026 (SWE-bench data). The 95–96% aggregator claims for DeepSeek V4-Pro, GPT-5.6 Sol and Claude Opus 5 appear on neither the official board, Epoch's own runs nor any vendor table read, so they are treated as unverified. Vendor-selected comparisons: Anthropic pits Opus 5.5 at xhigh effort against GPT-6 Astra at OpenAI-reported high effort, and Google's card compares 3.8 Flash with older rivals. On the tbench.ai board the OpenAI and Google figures match within 0.3 points and Opus 5.5 and Fable 5.1 within about two; Sonnet 5.5's 70.6% is the outlier. Effort settings: on the Sonnet 5.5 page, FrontierCode is 46.2% at Max effort but 52.1% at Xhigh, and a secondary report says xAI's own table compares Grok 4.7 at its highest effort with Grok 4.6 at a lower one. Safeguards: Anthropic counts safety interventions as AutomationBench failures, and Fable 5 and 5.1 scored zero on OSWorld 2.0 tasks where safeguards intervened (Claude Fable). Margins: Anthropic itself writes that "benchmark margins have become a less reliable guide to real-world differences" (Anthropic). The practical rule is to use vendor tables to build a shortlist, then decide with your own evaluation set built from your real workflows. Frameworks and Tool stack describe the tooling.

Usage splits in two: cheap Chinese models win volume, Claude heads one per-task spend list

OpenRouter is the only public, model-level usage source, and its data shows two markets. By requests, the four Chinese authors listed (DeepSeek, Z.ai, Qwen and Xiaomi) took 37.0% in the week beginning 28 September and Anthropic took 2.5% (OpenRouter rankings). By tokens over the trailing 30 days, the leaders were DeepSeek V4.1 Flash (101T), a stealth model called Space Bunny Alpha (56.7T), GLM 5.3 Flash (55.1T), Tencent's Hy4 preview (40.6T) and GPT-5.6 Luna (33.9T). No Claude model was in the 30-day top ten, although Opus 5.5 was tenth for the week at 4.68T. By spend, the order reverses in the one list captured. OpenRouter's "Top models by task" section ranks "each task's leading models ... by share of spend on OpenRouter", and the list read for this research was led by Opus 5.5, Sonnet 5.5 and Fable 5.1, followed by Qwen3.8 Max, GPT-6 Astra and GPT-6.1 Sol. The task behind that list was not recorded, and the section did not render on 11 October. OpenRouter says its token leaderboards and request chart do not measure "users or spend" (OpenRouter rankings). The volume data show price-sensitive bulk traffic going to cheap Flash-class models; the one spend list fits the idea that money goes to frontier models used for harder work, but cannot show any vendor's overall share of spend.

Enterprise share data is older and secondary. A mid-2025 Menlo Ventures survey of 150 technical leaders put Anthropic at 32% of enterprise LLM API usage, OpenAI at 25% and Google at 20% (BigDATAwire). Secondary summaries of a late-2025 Menlo survey give Anthropic 40% of spend, OpenAI 27% and Google 21%, unverified because Menlo's page could not be opened (Longterm Wiki). No 2026 edition was reachable.

Vendor-reported scale adds context, though each vendor defines it differently. Alphabet says Gemini models "process 22 billion API tokens per minute" and the Gemini app has 950 million monthly users (Alphabet SEC filing). Anthropic reported run-rate revenue of $14 billion in February 2026 (Anthropic Series G) and above $30 billion by 6 April, with over 1,000 business customers spending more than $1 million a year (Anthropic), and its Series H announcement (28 May 2026) says run-rate revenue "crossed $47 billion" (Anthropic Series H); about $65 billion at end-July was reported by Bloomberg (17 Aug 2026; headline only, Bloomberg via HN) and is not confirmed by Anthropic. SpaceX's S-1 shows the AI segment that includes Grok with $3.2 billion of 2025 revenue, a $6.4 billion operating loss and 117 million users of Grok's AI features (SpaceX S-1/A). India is the second-largest Claude.ai market at 5.8% of consumer use in Anthropic's February 2026 India brief (Anthropic India brief); the June 2026 Economic Index file puts India at 6.92% (April) and 7.12% (May 2026) of Claude.ai usage, with a usage-per-capita index of 0.31 and 0.30 (Anthropic Economic Index data), and ChatGPT has more than 100 million weekly users there (TechCrunch). Multi-vendor use is now normal: GitHub Copilot lists Claude, GPT, Gemini Flash, Grok and Microsoft's MAI-Code models side by side (GitHub Docs). Open-weight adoption shows up in downloads. On Hugging Face, gpt-oss-20b had 5.9 million, Qwen3.8-27B 6.8 million and GLM-5.3-Flash 6.54 million (Hugging Face; OpenAI gpt-oss).

Which model for which automation job

The matrix below turns the evidence into first choices. The cost figures come from token profiles in the research set's cost model. Those profiles are stated assumptions, not measurements, and should be replaced with logged data from a pilot (Costs).

Sources for the matrix: the vendor pricing pages cited above, plus IBM Granite, Phi-4-mini, Retell AI pricing and Lambda pricing.

The per-task arithmetic explains why the matrix splits work across tiers. Most of these numbers come from the cost model in this research set at 2026 list prices.

Note on the Haiku code-change cell: the prompt passes 100K tokens in the last 8 of 40 turns, where Haiku costs $0.50/$2.50, so the figure is about $0.28 rather than the flat-rate $0.136 (₹13.04). Long Haiku loops should be costed with that step.

Support products charge $0.99 per Intercom Fin outcome and $1.50 to $2.00 per Zendesk automated resolution (Intercom; Zendesk). A Sonnet-handled resolution costs about 5% of the Intercom fee in tokens, so when you compare a per-outcome product with building your own, engineering, retries, evaluation and support are the costs that matter. Voice is the exception, because platform minutes of $0.07 to $0.31 dominate the bill (Retell AI). For long documents, surcharges matter more than context size. Anthropic charges standard rates across the full 1M window on Claude 4.6 and later, except Haiku 5.5 (Anthropic pricing). For private data, the self-hosting break-even depends on throughput. One H100 at Lambda's $3.99 self-serve rate, which applies only to its largest H100 instance ($4.09–4.29 on smaller ones), beats a $0.17-per-million API price only above about 6,500 sustained output tokens per second at full utilisation (Lambda), and hosts already sell gpt-oss-120b at about $0.04 per million blended tokens (Artificial Analysis providers). In India, commercial H100 rental runs from about ₹217 to ₹362 an hour before 18% GST in one July–August snapshot, and E2E Networks' H100 page listed ₹255.55 an hour on demand and ₹70 spot on 11 October 2026 (ECorpIT; E2E Networks). Self-hosting is therefore usually a data-residency decision, not a cost decision. Hardware covers sizing.

Route cheap first and escalate on evidence

A router sends each request to the cheapest model likely to succeed and escalates only on a measurable signal. The tiers below map the top-ten models to that pattern.

The economics favour this design. Suppose 80% of support cases end at a Haiku 5.5 agent and 20% escalate to Sonnet 5.5, paying for both attempts. The blended cost per resolution is then about $0.013 (₹1.26), against $0.052 (₹5.01) for sending everything to Sonnet. That is a reduction of about three-quarters. This is illustrative arithmetic on the cost profiles above with an assumed escalation rate, and real escalation rates should come from a pilot.

Six levers matter as much as the routing rule. Caching: cache reads at 5% of input on Opus and Sonnet 5.5, and $0.003 per million on DeepSeek Flash, reward stable system prompts and tool definitions. Those definitions cost tokens too: Anthropic's computer-use toolset adds about 4,500 input tokens per request and the browser toolset about 6,600 (Anthropic pricing). Batch and off-peak: Anthropic, OpenAI and Google take about 50% off batch work; xAI's 20% batch discount does not cover Grok 4.7 (xAI pricing). DeepSeek's peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays (06:30–09:30 and 11:30–15:30 India time), so bulk DeepSeek jobs belong after 15:30 India time or at weekends (DeepSeek pricing). Effort: Opus 5.5 at high effort costs about a third of max effort for 4 fewer AA points, so default low and raise effort per task class. Version pinning: Google routed gemini-3.7-flash to 3.8 Flash on 8 October (Gemini API changelog), DeepSeek's Anthropic-compatible endpoint silently maps claude-opus names to V4-Pro and other Claude names to Flash (DeepSeek Anthropic API guide), and Anthropic released seven models between 9 June and 7 October (Claude release notes). Pin model IDs and re-run evaluations at every change. Budget guards: Anthropic's Start tier stops at a $500 monthly cap with HTTP 429 and no retry-after header until the next month, unless the account moves to a higher tier sooner (Anthropic rate limits), so unattended pipelines need spend alarms. Cross-vendor fallback: vendor-neutral plumbing keeps a second vendor cheap to hold on standby. MCP sits under the Linux Foundation's Agentic AI Foundation (AAIF), to which Anthropic (MCP), Block (goose) and OpenAI (AGENTS.md) contributed the founding projects (Linux Foundation; Anthropic), A2A is governed under the Linux Foundation (A2A), and xAI's API works with the OpenAI SDK. Frameworks compares the orchestration libraries that implement routers.

Residency, retention and licences decide where your data can go

For business and personal data, the shortlist is set by contracts as often as by capability. The controls found in the research:

Sources: Anthropic data residency, Anthropic regional compliance, Anthropic pricing, OpenAI pricing, OpenAI latest-model guide, OpenAI: Your data, TechCrunch, Google Cloud pricing, xAI regional endpoints, xAI security FAQ, Wikipedia: Grok deepfake scandal (secondary), Anthropic API and data retention, AWS Bedrock regional availability, AWS Bedrock: Claude Sonnet 5 model card, Azure Foundry deployment types, Azure Foundry region availability, Google Cloud data residency, DeepSeek privacy policy, DeepSeek Open Platform terms, NIST CAISI 2025, Kimi K3 licence, Qwen licence, MiniMax licence, Mistral pricing.

Two conclusions follow. For US businesses, the safe defaults are US-hosted frontier APIs with zero-retention or US-only options. China-hosted endpoints, including DeepSeek's API, the Kimi and Z.ai platforms and Alibaba's Beijing region, should not receive customer or business data, and Xiaomi's MiMo API should be treated the same until its terms are read. For Indian SMEs, no vendor's own API processes requests in India: OpenAI's India region stores data at rest but processes it elsewhere. AWS Bedrock does offer India-only routing (Mumbai and Hyderabad) for the previous Claude generation (Opus 5, Sonnet 5, Haiku 4.5), Kimi K3 and GPT-5.6 Terra/Luna, and lists older open-weight models In-Region in Mumbai. The newest models (Claude 5.5, GPT-6, Gemini 3.6 and later) have no documented India processing on any cloud read. The options are Bedrock's India profiles with a previous-generation model; OpenAI's India storage option or a US or global endpoint under a data-processing agreement; open models on Indian clouds (E2E Networks' TIR endpoints, Yotta Shakti Cloud endpoints; vendor-reported) (E2E Networks; Yotta); or self-hosted open weights on Indian GPU capacity. India's DPDP Rules phase in from 13 November 2026 (consent managers) to 13 May 2027 (the main notice, consent, security and breach duties). This document did not analyse how they apply to model APIs, so check that with counsel. Governance and India cover the regulatory side.

What this means for your business

Shortlist by job and tier, not by vendor. On this page's evidence the frontier tier is Opus 5.5, Sonnet 5.5, GPT-6 Astra and GPT-6.1 Sol; the value tier is Haiku 5.5, GPT-6 Luna, Gemini 3.8 Flash and, for non-sensitive data, DeepSeek V4.1 Flash; and gpt-oss, Qwen3.8-27B and GLM-5.3-Flash are open-weight paths for private data. Keep two frontier vendors on the shortlist as a guard against refusals, outages, gating and price changes. Astra is in a staged release, and Anthropic says a US export-control directive on 12 June 2026 suspended foreign nationals' access to Fable 5 and Mythos 5, after which it disabled both models for every customer (Anthropic statement; Al Jazeera); access was later restored (Claude release notes).

If data must be processed in India, the shortlist changes. Claude 5.5, GPT-6 and Gemini 3.6 and later have no documented India processing. The documented options are previous-generation models on Bedrock's India profiles (Claude Opus 5, Sonnet 5 or Haiku 4.5, GPT-5.6 Terra or Luna, Kimi K3), Gemini 3.5 Flash or 2.5 Flash in Mumbai, open models on Indian clouds, or self-hosted open weights.

Budget the work, not just the tokens. On the cost profiles above, model tokens are 0.3% to 5% of a $0.99 per-resolution reference price, and a value-tier invoice extraction costs about seven paise. Engineering, retries, evaluation and support make up most of the cost. Routing, caching, batching and lower effort settings cut tokens further, and unattended pipelines need spend alarms.

Diary the deadlines. OpenAI's Agent Builder shuts down on 30 November 2026 (OpenAI), and Claude Sonnet 4.5 retires from the Claude API the same day. OpenAI's deprecations page lists further shutdowns on 23 October and 11 December 2026 (including the original GPT-5 snapshots and o3) and 1 April 2027, and no new fine-tuning jobs from 6 January 2027 (OpenAI deprecations). GPT-5.6 Sol's promotional price runs "at least through" 21 November, Gemini 3.8 Flash's price doubles on 1 January 2027, and India's main DPDP duties apply from 13 May 2027.

Test before you commit, and keep testing. Vendor benchmark tables are self-selected, the independent agentic board does not reproduce some vendor figures, and METR's autonomy measure broke down on cheating. Gartner forecasts that more than 40% of agentic AI projects will be cancelled by the end of 2027; the title was confirmed, but the body was not readable (Gartner). Build an evaluation set from your own workflows, re-run it at every model change, and check coding agents with tests they cannot see (Playbook).

Keep a vendor risk register. xAI's AI segment loses money, and Grok faces open regulatory proceedings in the EU and UK and unresolved MeitY notices in India. Chinese APIs raise data-path questions, and Chinese open-weight licences add revenue thresholds that matter mainly if you offer hosted models as a service. None of this rules a vendor out (Landscape; Companies; Use cases).

Questions for any vendor or implementation partner: Which model IDs will run, and what happens when they retire? Does the contract name model classes and say how list-price changes pass through? Where is each request processed, and which features fall outside zero data retention? Which second-vendor model is the fallback? How much of the price is tokens, and how much is engineering and support?

Open questions and unverified items

Conflicts resolved in this document: Grok 4.6's "AA 61" from a news site is Artificial Analysis's speed figure, and AA's own index is 44 (Artificial Analysis). Opus 5.5's "58" on the AA leaderboard and "57.6" in OpenRouter's record are the same value rounded (the leaderboard's embedded data gives 57.6). An earlier finding that OpenRouter had no GPT-6.1 Sol record was wrong: the record exists at 51.8. Kimi K3 and GLM 5.3 prices: the vendor pages give the list prices, and OpenRouter figures are dated host listings. CAISI's $1.74/$3.48 price for DeepSeek V4-Pro was the April price, superseded by $0.66/$1.98 off-peak. Earlier references to Terminal-Bench 2.0 and OSWorld-Verified are superseded by versions 4.0 and 2.1. Google's own announcement confirms Gemini 4 Argon, replacing an earlier "not confirmed" note, but it does not confirm general availability. The Timeline records release dates, and the index lists all 26 documents in the set.

Sources

1. OpenRouter models API
2. OpenRouter rankings
3. Artificial Analysis model leaderboard
4. Artificial Analysis: Grok 4.7
5. Artificial Analysis: Grok 4.6
6. Artificial Analysis: gpt-oss-120b providers
7. Artificial Analysis Intelligence Index v4.2 note (yorozuipsc, third-party)
8. Epoch AI benchmarks hub
9. Epoch AI Capabilities Index
10. Epoch AI data insights
11. Epoch AI trends
12. Epoch AI benchmark data (bulk download)
13. arena.ai leaderboard
14. LMArena text leaderboard (8 Oct 2026 snapshot)
15. LMArena WebDev leaderboard (8 Oct 2026 snapshot)
16. SWE-bench official leaderboard data
17. Scale: SWE-bench Pro leaderboard
18. Terminal-Bench
19. OSWorld 2.0 paper (arXiv 2606.29537)
20. METR time horizons
21. METR: Claude Opus 5.5 pre-deployment summary
22. METR: GPT-5.6 Sol evaluation
23. METR: Time Horizon 1.1
24. NIST CAISI evaluation of DeepSeek V4 Pro
25. NIST CAISI evaluation of DeepSeek models (Sept 2025)
26. NIST CAISSI (formerly CAISI) home page
27. Anthropic: Claude Opus 5.5
28. Anthropic: Claude Sonnet 5.5
29. Anthropic: Claude Haiku 5.5
30. Anthropic: Claude Fable
31. Anthropic: statement on the directive to suspend Fable 5 and Mythos 5 access
32. Al Jazeera: US lifts restrictions on Fable and Mythos models (1 Jul 2026)
33. Claude Platform pricing
34. Claude models overview
35. Claude data residency
36. Claude API and data retention
37. Claude Platform release notes
38. Claude rate limits
39. Claude extended thinking
40. Claude regional compliance
41. Anthropic Series G announcement
42. Anthropic: Google and Broadcom compute partnership
43. Anthropic: Series H announcement
44. Anthropic: Economic Index India brief
45. Anthropic Economic Index dataset (Hugging Face)
46. Anthropic: donating MCP and establishing the Agentic AI Foundation
47. Linux Foundation: formation of the Agentic AI Foundation
48. OpenAI API pricing
49. OpenAI: GPT-6 Astra model page
50. OpenAI: Using GPT-6
51. OpenAI: Your data (data retention and residency controls)
52. OpenAI: reasoning models
53. OpenAI: Agent Builder
54. OpenAI: deprecations
55. OpenAI: GPT-6 Astra system card
56. OpenAI Deployment Safety Hub
57. Constellation Research: OpenAI launches GPT-6 Astra
58. TechCrunch: OpenAI taps Tata for India data-centre capacity
59. Gemini API pricing
60. Gemini API changelog
61. Google Cloud: Gemini pricing on Agent Platform
62. Google Cloud: data residency on Agent Platform
63. Google: Gemini 4 Argon
64. Google DeepMind: Gemini 3.8 Flash model card
65. Alphabet Q2 2026 results (SEC exhibit 99.1)
66. xAI: models and pricing
67. xAI: pricing
68. xAI: regional endpoints
69. xAI: structured outputs
70. xAI: security FAQ
71. SpaceX S-1/A amendment 2
72. Wikipedia: Grok sexual deepfake scandal (secondary)
73. Business Today: MeitY notice to X over Grok (2 Jan 2026)
74. Storyboard18: X removes Grok content after MeitY notice
75. DeepSeek API pricing
76. DeepSeek API changelog
77. DeepSeek Anthropic API guide
78. DeepSeek privacy policy
79. DeepSeek Open Platform terms of service
80. Hugging Face: DeepSeek-V4.1-Flash
81. Hugging Face API: DeepSeek-V4.1-Flash
82. AWS Bedrock: model support by Region
83. AWS Bedrock: Claude Sonnet 5 model card
84. Microsoft Foundry: deployment types
85. Microsoft Foundry: region availability for models sold by Azure
86. Yotta Labs: DeepSeek V4 Flash hardware requirements
87. Alibaba Cloud Model Studio pricing
88. Qwen3.8-2.4T-A95B licence
89. Hugging Face: Qwen3.8-27B
90. Kimi K3 licence
91. Kimi platform pricing
92. Z.ai pricing
93. Hugging Face API: zai-org models
94. Hugging Face API: XiaomiMiMo/MiMo-V2.6-Pro-RL
95. Hugging Face: XiaomiMiMo/MiMo-V2.6-Pro-RL repository
96. Xiaomi MiMo pay-as-you-go pricing
97. MiniMax pay-as-you-go pricing
98. MiniMax M3 licence
99. Hugging Face: gpt-oss-120b
100. Mistral: Introducing Mistral Large 4
101. Mistral pricing
102. IBM Granite
103. Hugging Face: Phi-4-mini-instruct
104. GitHub Docs: supported AI models in Copilot
105. BigDATAwire on the Menlo Ventures mid-2025 report
106. Longterm Wiki sourcing page on Menlo late-2025 figures (secondary)
107. Intercom pricing
108. Zendesk pricing
109. Retell AI pricing
110. Lambda pricing
111. ECorpIT: India GPU cloud rental pricing
112. E2E Networks: H100 pricing in India
113. E2E Networks: inference endpoints
114. Yotta Shakti Cloud: AI endpoints
115. A2A protocol
116. Gartner: over 40% of agentic AI projects will be canceled by end of 2027
117. e-Gazette: DPDP Rules commencement notification G.S.R. 843(E)
118. X-Rates: USD to INR monthly averages, 2026
119. Bloomberg headline on Anthropic run-rate, 17 Aug 2026 (Hacker News listing; headline only)