Costs and Economics of AI Automation

Key takeaways

• Model tokens are the cheap part. A support reply costs about $0.0003 (₹0.03) on Haiku 5.5 or GPT-6 Luna and $0.006 (₹0.60) on Sonnet 5.5. Across the models tabled below, a five-turn resolution agent costs $0.003 to $0.14 and a 40-turn coding agent $0.14 to $7.49, rising to $11.80 to $13.58 on Fable 5.1 or GPT-6 Astra (assumed token profiles; arithmetic below). Published per-outcome prices of $0.99 to $2.00 are 19x to 38x the Sonnet 5.5 cost of one resolution.
• List prices span about 300x, from $0.17 per million output tokens (gpt-oss-120b) to $50 (Claude Fable 5.1, GPT-6 Astra). Budget per model tier, and send easy steps to cheap tiers.
• Two price trends run at once. Fixed capability gets about 10x cheaper a year or faster, while frontier prices hold or rise. Gemini 3.8 Flash doubles on 1 January 2027, and GPT-5.6 Sol's promotional price is guaranteed only through 21 November 2026.
• Hidden drivers often move the bill more than the headline price. These include caching (reads at 2% to 10% of input), long-prompt surcharges (1.5x to 5x), web search ($10 to $14 per 1,000 calls), fast modes (2x to 6x) and data-residency uplifts (10%). Parsing can cost 5x to 160x the model in document work, and per-minute platform fees dominate voice.
• Self-hosting rarely beats the API at SME volumes. A rented H100 at Lambda's $3.99 an hour (its largest-instance rate; smaller instances cost $4.09 to $4.29) breaks even with a $0.17-per-million API only at about 6,500 tokens per second around the clock. A DGX Spark measured at 353 tokens per second matches the API only at full use and before staff time.
• Product anchors: $17 to $40 per seat per month for assistants; $0.99 to $2.00 per resolution or $0.005 to $0.008 per credit for agents.
• Implementation partners' published fees come in three steps. US boutiques list an assessment for $3,500 to $8,500, a build sprint for $18,000 to $60,000, and managed service from $1,500 a month (₹3.4 to 8.2 lakh, ₹17.3 to 57.6 lakh, ₹1.44 lakh a month). Indian firms rarely publish prices.
• Capex runs far ahead of AI revenue. Big-four platform capex rose about 80% year on year, and Microsoft's AI run-rate is about a third of its capex. The Bank of England and the IMF warned of a correction. Check how your vendor and partner contracts handle price reviews and exchange rates.

How to read the numbers

All facts are stated as of 12 October 2026. Model list prices, reasoning-token rules, GPU rates and Indian compute and messaging prices were re-checked against primary sources on 12 October 2026; facts read only on 11 October keep that read date. Prices are in US dollars per million tokens (input / cached input / output) unless marked otherwise. INR conversions use ₹96 per US dollar, the X-Rates monthly average for October 2026 (95.95 over the first 11 days, rounded) (X-Rates monthly average). One lakh is ₹100,000.

The source labels used throughout are as follows. Vendor means the figure was read on the vendor's own pricing or documentation page. Vendor (summary) means the vendor page was read through a summarising reader, so spot-check it before quoting. OpenRouter means an OpenRouter list price only, which is provider-specific and not first-party. Secondary means a blog, news report or Wikipedia. Vendor claim marks a vendor's statement about its own product. Assumption marks an input I chose. Arithmetic marks my calculation from stated figures. Model details are in Top models, Claude, OpenAI, Gemini, Grok, DeepSeek and Other models. Hardware is covered in Hardware and supporting tools in Tool stack.

List prices span a 300x range, so budget by tier rather than by vendor

The master table below lists 23 models that a 2026 automation buyer is likely to consider, sorted by output price. Prices for Anthropic, OpenAI, xAI and DeepSeek were confirmed on vendor pages (re-read 11–12 October 2026), and Moonshot, Z.ai, MiniMax and Alibaba list prices on their own pages on 12 October. The OpenRouter models API matched Anthropic, OpenAI, xAI and DeepSeek on every row checked with two exceptions. It lists GPT-5.6 Sol at $2 / $10, which is OpenAI's Batch price. For DeepSeek it carries the off-peak tier for V4-Pro but the peak tier for V4.1 Flash (OpenRouter models API). For Kimi K3 and GLM 5.3, OpenRouter's host listings sit well below the vendors' own list prices and vary between reads. Rows marked OpenRouter have no first-party confirmation.

Sources for the table: Anthropic (Claude Platform pricing; Opus 5.5; Haiku 5.5), OpenAI (OpenAI API pricing; OpenAI pricing, markdown; GPT-6 Astra model page), Google (Gemini API pricing; Google Cloud Agent Platform pricing), xAI (xAI models; xAI API pricing), DeepSeek (DeepSeek API pricing), Moonshot (Kimi pricing), Z.ai (Z.ai pricing), MiniMax (MiniMax pay-as-you-go), Alibaba (Model Studio pricing), and OpenRouter for the rest.

The spread is the first fact to plan around. Output prices run from $0.17 to $50, a ratio of about 300x (arithmetic), and similar-looking "mini" or "flash" tiers differ by 7.5x on input ($0.75 for Gemini 3.8 Flash against $0.10 for GPT-6 Luna and Haiku 5.5). Usage follows price. In the week beginning 28 September 2026, DeepSeek models carried 21.8% of OpenRouter requests, Google 21.5% and OpenAI 18.2%, against 2.5% for Anthropic. OpenRouter's "Top models by task" section is ranked by share of spend, and Claude models headed the one per-task share-of-spend list captured in the research (its task was not recorded). That is not an overall spend ranking: OpenRouter says its token leaderboards and its request-share chart, the source of the shares above, do not measure users or spend (OpenRouter rankings). The request data shows high-volume steps flowing to cheap models. It cannot show where the money goes.

Several rows need care before they go into a budget. GPT-5.6 Sol is listed by OpenAI at $4 / $20 under a note that its "promotional pricing is available at least through November 21, 2026", with no post-promotion price stated. OpenRouter lists it at $2 / $10, which is the Batch price on OpenAI's page. No post-promotion price is assumed here; I use OpenAI's own page and flag the figure as promotional (OpenAI API pricing). GPT-5.5 ($5 / $0.50 / $30 below 272K tokens, $10 / $1 / $45 above) and the other GPT-5.x rows, such as GPT-5.6 Terra at $2 / $0.20 / $12, are official OpenAI prices. They appear in the full tables of OpenAI's pricing page, re-read on 11 October 2026 (OpenAI pricing, markdown). The HTML page's headline tables show only the newest models, which is why earlier reads treated these rows as unconfirmed. OpenRouter also lists GPT-5.5 at $5 / $30. The DeepSeek V4-Pro official price of $0.66 / $1.98 off-peak is lower than the $1.74 / $3.48 that a report by NIST's CAISI (which NIST's site now calls CAISSI) used (NIST CAISSI). Those were the developer-reported prices in force during CAISI's April 2026 test, published on 1 May 2026 (NIST CAISI evaluation). The current official page supersedes them. The likely cause is a repricing when V4-Pro reached general availability on 13 August 2026, but that is an inference. GLM 5.3's own list price is $1.40 / $4.40, about 3:1 (arithmetic); the 69:1 ratio appears only in OpenRouter's host listing ($0.07 / $4.80), so check the host before planning reading-heavy work on that price (Z.ai pricing). For open-weight models, the price depends on the host. Artificial Analysis lists 17 providers for gpt-oss-120b, whose prices differ by up to 8.9x, with the cheapest blended price at $0.04 per million (CoreWeave) (Artificial Analysis providers).

Capability gets cheaper every year while the frontier price holds

The long-run trend is steep deflation for a fixed level of capability. a16z found the cheapest model scoring 42 on MMLU fell from $60 per million tokens in November 2021 to $0.06 in late 2024, "a factor of 1,000 in 3 years", and called a 10x-a-year decline "a fairly good approximation" (a16z, LLMflation). Epoch AI measured declines of 9x to 900x a year depending on the benchmark, and 40x a year for GPT-4-level PhD-science questions. It cautions that the fastest drops were recent and may not persist (Epoch AI price trends). On the hardware side, Epoch puts AI-chip price-performance gains at about 49% a year since 2023 (Epoch AI trends). No 2026 update of Epoch's API price series was found.

The frontier tells a different story. OpenAI's GPT-5 launched at $1.25 / $10 in August 2025 (launch date per OpenRouter's record), while GPT-5.5 lists at $5 / $30 and GPT-5.5 Pro at $30 / $180 (OpenAI pricing, markdown). Google moved from Gemini 2.5 Pro at $1.25 / $10 to Gemini 3.1 Pro Preview at $2 / $12 (Gemini API pricing). Anthropic went the other way at the top. Opus 4.1 cost $15 / $75, Opus 5 cost $5 / $25, and Opus 5.5 costs $4 / $20, "20% less than Opus 5" (Opus 5.5). Anthropic then added a tier above Opus, with Fable 5.1 at $10 / $50 (Claude Platform pricing). Anthropic also says Opus 5.5 is about 40% cheaper per task than Opus 5 at default settings (vendor claim). That is a reminder that price per token and cost per task differ when models use different numbers of tokens. Compute input costs have risen too. An H100 rental contract index rose from $1.70 to $2.35 per GPU-hour between October 2025 and March 2026, about +38% by arithmetic (single secondary source, whose body says about 40% and headline 50%) (Crypto Briefing).

The working rule for budgets follows from both trends. Assume a price cut for "the same quality next year", and a flat or higher price for "the best model available". Then check the dated changes below, which have known effective dates.

Hidden cost drivers can double a token estimate

A list-price estimate is a floor. Caching, reasoning, long prompts, tool calls, speed tiers and routing each change the bill, and they compound in agent loops that re-send a growing context.

No published study gives a measured overhead factor for production LLM workloads. The research applies a multiplier of 1.2x to 2x to raw model cost to cover retries, evaluation runs and monitoring traffic. That multiplier is an assumption, not a sourced figure, and it should be replaced with logged data after a pilot.

Cost per task: fractions of a cent for text, dollars for agents

Every task cost below uses one formula. Cost equals uncached input tokens times the input rate, plus cached tokens times the cache rate, plus output tokens times the output rate, all divided by one million, plus any tool fees. The token profiles are assumptions chosen to represent typical flows, not measurements. Cache writes are ignored (under about 5% at these sizes), and prices are standard tier with no batch discount. One full example shows the method. A five-turn support resolution on Sonnet 5.5 is (18,000 × $2 + 12,000 × $0.10 + 1,500 × $10) ÷ 1,000,000 = ($36,000 + $1,200 + $15,000) ÷ 1,000,000 = $0.0522, or ₹5.01.

The computed table is from the primary economics research, re-checked by script. Its inputs are the vendor prices in the master table and the tool fees in the previous section.

Four points stand out. First, text automation is nearly free at the model layer. Ten thousand Sonnet 5.5 support replies cost $62 a month, or $31 with batch (arithmetic: 10,000 × $0.0062). A thousand invoices on Haiku 5.5 cost $0.75. Second, search fees dominate lead research on cheap models. In T5 on Haiku 5.5, the $0.03 of searches is 89% of the $0.0337 total (arithmetic: (24,000 × $0.10 + 12,000 × $0.01 + 2,400 × $0.50) ÷ 1,000,000 = $0.0037 of tokens). For DeepSeek, the search fee assumes a third-party search tool at the same $10 per 1,000. Third, agent loops are where model choice matters. The 40-turn coding task costs about $1 on Gemini 3.8 Flash and $5 to $7.50 on Opus 5.5 or GPT-5.5. The same task on GPT-6 Astra or Fable 5.1 costs $13.58 or $11.80 (arithmetic; Fable's 0.025x cache rate makes it cheaper than Astra on this cache-heavy profile). Fourth, Haiku 5.5's 100K price step matters in long loops. At base rates T6 would cost $0.136 on Haiku 5.5. But the profile's context grows from 20K to 120K tokens, so its last eight turns pass 100K, and each of those whole requests is billed at $0.50 / $2.50. That raises the task to about $0.28 (₹26) (arithmetic, assuming the context grows evenly). Any base-rate Haiku estimate for a long agent loop is therefore a lower bound. GPT-6 Luna's long-context rate starts only above 272K tokens, so $0.136 stands for Luna.

Document automation shows how non-model costs take over. A scanned 3-page invoice costs $0.045 to parse on Unstructured, $0.03 to $0.12 on Reducto, and about $0.004 on LlamaParse basic parsing. Against $0.00075 of Haiku tokens, that is 5x to 160x the model cost (arithmetic). Scanned-page image tokens were not priced in the research. The table below converts monthly volumes to rupees for an Indian SME budget (arithmetic at ₹96, model cost only).

The research found no measured production token profiles for any of these tasks, so all six profiles should be replaced with logged data from a pilot. Use-case evidence on returns is in Use cases.

Voice and messaging are priced per minute and per message

Voice is the exception to "tokens are cheap", because the platform charges per minute whatever the model costs. Retell publishes $0.07 to $0.31 per minute for AI voice agents. Its worked example of $0.11 a minute breaks down into $0.04 for the LLM, $0.055 for voice infrastructure and $0.015 for text-to-speech, with no telephony charge in that example (Retell AI pricing). At those rates an hour of calls costs $4.20 to $18.60, and a five-minute call $0.35 to $1.55, or ₹34 to ₹149 (arithmetic).

A voice budget should be built as platform plus LLM plus speech-to-text plus text-to-speech plus telephony. Retell's $0.11 and LiveKit's $0.0479 examples show how far the total moves with component choice, and neither includes an Indian telephony leg. Plivo's India page lists ₹0.38 and ₹0.25 a minute (rows not labelled by direction) and ₹200 a month for a domestic number (Plivo India voice pricing). Bolna's $15 pilot plan is ₹5.52 (6 cents) a minute (Bolna pricing). Exotel's own blog, reported through a search summary (not read directly), gives about ₹0.80–1.00 a minute outbound and ₹0.30–0.50 inbound (single-source; Exotel blog); Twilio India and Exotel rate cards were not read (all vendor-reported). TRAI's draft Third Amendment to the TCCCPR (comments April 2026) adds an A2P call definition with pre-declaration and termination charges; a fully automated voice agent appears to fall within it (inference) (comments hosted by TRAI). Final rules were reported as notified on 18 Sep 2026 (secondary); the official text was not read. For Hindi and other Indian languages, Sarvam's INR pricing removes exchange-rate exposure on the speech leg. Its quality against English-first vendors was not benchmarked.

WhatsApp is the main messaging channel for Indian SMEs, and its costing is unsettled. Meta's page confirms a higher India marketing rate from 1 January 2026 and a higher authentication-international rate from 1 April 2026. Businesses billed in India must move every WhatsApp Business account to INR billing by 31 December 2026: "as of January 1, 2027 Meta will no longer deliver the messages of non-INR WABAs" (Meta WhatsApp pricing, read through a summary). Per-message India rates from 1 October 2026, excluding 18% GST, are vendor-reported (ChatMaxima; Meta's rate-card file not read): marketing ₹0.8631, utility ₹0.1150, authentication ₹0.1150, authentication-international ₹2.4971. A blog gives the same marketing rate, up from ₹0.7846 (Authkey). Gallabox's INR tab shows exactly 1.15 times those (marketing ₹0.9926, others ₹0.1323), reason not stated (Gallabox). Two vendor pages say service messages from 1 October 2026 are free up to 1,000 per business phone number a month, then billed at the utility rate (TelecomTalk's ₹0.115 matches) (TelecomTalk); Meta's own page, as read on 11 October, still said non-template messages are free (vendor-reported; not confirmed on Meta's page). Per-message INR costs should be confirmed against Meta's current rate card.

Self-hosting beats the API only at sustained, high utilisation

The self-host question is a throughput question. A rented GPU costs the same whether it is busy or idle. The API market already prices open models near cost, so a small business wins on price only if it keeps hardware busy around the clock. Hardware options and sizing rules are in Hardware.

Worked example A: one rented H100 against the gpt-oss-120b API. Lambda's self-serve H100 SXM costs $3.99 per GPU-hour on its largest (208-vCPU) instance, and $4.09 to $4.29 on smaller configurations (re-read 12 October 2026) (Lambda pricing). The steps use $3.99:

1. Monthly cost = $3.99 × 730 hours = $2,912.70 (₹2.80 lakh).
2. Break-even volume against the $0.17-per-million OpenRouter output price = $2,912.70 ÷ $0.17 = 17,134 million, or about 17.1 billion output tokens a month.
3. A month has 2,628,000 seconds, so that volume needs about 6,520 tokens per second at 100% utilisation, or 13,040 at 50%.
4. Against the $0.04 blended price from the cheapest provider, break-even rises to 72.8 billion tokens a month, or about 27,700 tokens per second.

At the $4.29 smaller-instance rate, the month costs $3,131.70 and break-even rises to about 7,000 tokens per second (arithmetic). This comparison uses the output price only. Adding the API's input charges would lower the break-even somewhat. Lambda's reserved 16-GPU cluster rate of $6.16 per GPU-hour is higher than its self-serve rate, a difference the page did not explain. The table below shows the effective GPU cost per million output tokens at different sustained throughputs (primary research, arithmetic).

MLPerf Inference v6.1 (vendor submissions under MLPerf rules) has no H100 rows, but it measured gpt-oss-120b in the Offline scenario at about 11,400 tokens/s per B200 and 1,940–1,970 per RTX PRO 6000 (8-GPU servers) (MLPerf Inference v6.1 results). At that rate a B200 rented at E2E's $6.99 an hour ($5,103 a month) breaks even with $0.17 per million output tokens only at about 100% utilisation (arithmetic: $5,103 ÷ $0.17 = 30.0 billion tokens a month, against 11,400 × 2,628,000 = 30.0 billion at full use). For H100 the table rows remain thresholds to test. Hyperscaler rates make the case harder. Azure's 8-GPU H100 virtual machine lists at $98.32 an hour in East US, or $12.29 per GPU-hour, and a single-GPU H100 VM at $6.98 an hour (Azure retail prices API). AWS lists an H100 (p5) at $6.88 per GPU-hour in the US and $8.26 in Mumbai, which moves break-even to about 11,200 and 13,500 tokens/s (arithmetic: $6.88 × 730 = $5,022 a month, or 29.5 billion tokens at $0.17 per million, over 2,628,000 seconds; AWS Price List, Mumbai). Google Cloud lists committed-use discounts of 28% for one year and 46% for three years (Google Cloud VM pricing). The 2026 market range for on-demand H100s is $2 to $12 per GPU-hour, with a median of about $3.37 (secondary aggregator summary) (IntuitionLabs).

Worked example B: renting in India. E2E Networks' pricing page shows an H100 at $2.69 an hour in its USD table; the page title's "Starting ₹49/hr" is unexplained and is not used (E2E Networks pricing). E2E's H100 pages give ₹255.55 an hour on-demand, exactly the $2.69 USD price at E2E's fixed ₹95 per dollar (a vendor conversion, not FX), and ₹70 spot; a 3-month-plus commitment is "as low as ₹155.90" per GPU-hour (E2E H100 pricing in India; E2E H100 product page). A third-party snapshot from July to August 2026 puts domestic commercial H100s at ₹217 to ₹362 an hour, excluding 18% GST (ECorpIT). Earlier reads of the explainer showed ₹350 to ₹400, so check a live portal before you budget. At $2.69 an hour, one H100 costs $1,963.70 a month (₹1.89 lakh), and break-even against $0.17 per million needs 11.6 billion tokens a month, or about 4,400 tokens per second at full utilisation (arithmetic). At ₹217 to ₹362 an hour, the month costs ₹1.58 to 2.64 lakh before GST, or ₹1.87 to 3.12 lakh with it (arithmetic). IndiaAI's compute portal lists H100 SXM at ₹153 per GPU-hour before subsidy (GST extra), with a subsidy of "Max upto 40% subject to approval from IndiaAI" (₹91.80 at the maximum) (IndiaAI price calculator). At ₹153 one H100 costs ₹1,11,690 a month ($1,163), and break-even against $0.17 per million needs 6.8 billion tokens a month, or about 2,600 tokens per second; at ₹91.80 it is ₹67,014 ($698), 4.1 billion tokens and about 1,560 tokens per second (arithmetic, 730 hours, before GST). A minister's "about ₹65" (Storyboard18), and a secondary ₹67, cannot be reproduced from the portal. Access is by application: researchers and academia, DPIIT-recognised startups and MSMEs meeting revenue or funding tests, fellows, students and government entities; other companies, large enterprises included, are not a named category (inference). Requests under 5,000 GPU-hours are auto-approved (IndiaAI eligibility; see India).

Worked example C: an owned DGX Spark, using measured throughput. This is the one desk-side self-host case with a measured benchmark (MLPerf v6.1 covers data-centre servers; see above). llama.cpp's own benchmark file reports gpt-oss-120b (MXFP4) on a DGX Spark at 58.7 tokens per second for a single stream, falling to 42.8 at 32K tokens of context, and 353.3 tokens per second in aggregate across 32 parallel streams at a 512-token depth (llama.cpp DGX Spark benchmarks). One August 2026 report on an AI-assisted news site lists the box at $4,699 (secondary; NVIDIA's own page shows no price), up from a $3,999 launch price that rests on a search snippet (ainauten). Indian reseller listings show about ₹4.9 to 5.4 lakh including GST (search snippets; Variety Infotech). The steps are:

1. Maximum monthly output = 353.3 × 2,628,000 = 928 million tokens, at 100% utilisation and short context.
2. Hardware cost over 36 months (assumption) = ₹5,42,499 ÷ 36 = ₹15,069 a month, or $130.53 a month at the US price. That is about $0.14 per million tokens at full use, or $0.28 at 50%.
3. Staff time at 0.25 of a full-time engineer (assumption) paid the ₹10 lakh-a-year Glassdoor average for Indian AI engineers (Tredence, secondary) = ₹20,833 a month, salary only.
4. Total self-host cost ≈ ₹35,900 a month, before power (not sourced).
5. The same 928 million output tokens through the API cost 928 × $0.17 = $157.84, or ₹15,153 a month.

Even at a utilisation few teams reach, the API costs about 42% of the self-host option once staff time is counted (arithmetic). The 353 tokens per second figure is also at a short context. Long agent prompts will cut it.

A 96 GB RTX PRO 6000 workstation sits between these cases. The card launched at $8,565; NVIDIA's own Marketplace listed it at $16,000 by August 2026; tracked street prices were about $11,100–11,700 in June–July 2026 (Newegg $12,100 in June) (all secondary) (Tech Insider; Let's Data Science; GPU Poet). With an assumed $3,000 to $5,000 host, 36-month write-off, 0.7 kW draw and $0.12 per kWh (all assumptions), at mid-2026 street-to-Marketplace card prices it costs about $453 to $645 a month before staff (arithmetic: $14,100 ÷ 36 + $61 of power, up to $21,000 ÷ 36 + $61). MLPerf v6.1 measured gpt-oss-120b at about 1,940–1,970 output tokens/s per RTX PRO 6000 (Offline, vendor-tuned servers). At full use that is about 5.1 billion tokens a month, worth about $871 at $0.17 per million, so owning breaks even at roughly 52–74% sustained utilisation before staff time (arithmetic).

Self-hosting is justified by reasons other than price per token. The first is data that must not leave your control. India's DPDP Act obligations on notice, consent, security and breach reporting commence on 13 May 2027 (DPDP commencement notification; see Governance). The second is a fine-tuned model that no provider serves. That case is growing, because OpenAI says it "is winding down the fine-tuning platform", which is "no longer accessible to new users" (OpenAI API pricing). The third is sustained, predictable load. For an SME sending a few hundred thousand tokens a day, the API is cheaper than the staff time alone.

AI products price by seat, by outcome and by credit

Three pricing units now coexist. Seats remain the default for assistants and copilots. Outcomes (a resolved ticket) are becoming the default for customer-service agents. Credits, which meter agent steps, cover platform agents. The table gives the reference points to compare any quote against.

The per-outcome anchors are the most useful benchmark when you compare quotes for a support agent. Intercom's $0.99 and Zendesk's $1.50 to $2.00 set a reference band of roughly $1 to $2 per AI-resolved ticket (₹95 to ₹192). Against the $0.052 Sonnet 5.5 cost of one five-turn resolution (task T2), Intercom's fee is about 19x the model cost and Zendesk's 29x to 38x (arithmetic).

Anthropic's agent-design guidance notes that several companies offering customer-support agents charge "only for successful resolutions" (Anthropic, Building effective agents). Salesforce, which completed its roughly $3.6 billion acquisition of Fin (formerly Intercom) on 10 September 2026, says Fin agents resolve on average 76% of support volume end to end, across 30,000+ companies. Fin's own site gives the same 76% average across "12,000+ businesses" (fin.ai); the two vendor counts may measure different things (both vendor-reported; no definition of a resolution) (Salesforce Fin agreement; Salesforce Fin completion). Credit systems make cost depend on how many steps an agent takes, so the same agent-loop modelling used for tokens applies to credits. Seat prices cluster at $20 to $30 a month, with premium or heavy-use seats from $100. That cluster is the benchmark to measure any "AI employee" price against. Vendor profiles are in Companies.

What implementation partners charge: a three-step ladder, and Indian firms rarely publish prices

If you hire an implementation partner, published US boutique pricing shows what to expect. A fixed-fee assessment comes first, then a fixed-scope build, then a monthly retainer, with model usage passed through to you. Main & Machine budgets model usage and third-party subscriptions outside its fees (Main & Machine pricing). Softomate quotes "monthly care, licences, hosting and usage" separately (Softomate). Indian AI-automation firms surveyed in the research publish almost no package prices. Their pages say costs "vary based on the complexity of processes" (Carmatec) or depend "on workflow complexity and hosting preference" (Infyom). One says clients "pay for outcomes shipped rather than hours logged", without a price (Antino). Team and delivery detail is in Playbook.

No sourced Indian hourly rate card was found, so compare a partner's quote with the cost of doing the work in-house. A two-to-four-week assessment at the US band of ₹3.36 to 8.16 lakh compares with ₹41,667 to ₹83,333 of salary for one in-house Indian AI engineer at the ₹10 lakh average over the same period (arithmetic, salary only). A partner engagement usually involves more people and overhead than one engineer, so the comparison is not like for like. For most Indian SMEs, a US-priced assessment of ₹3.4 lakh or more is out of reach, so a smaller first step, such as a fixed-fee single-workflow pilot, fits better. That is an inference, as no Indian SME rate card was found.

Budgeting an automation project: what you pay for, and fixed-fee, value or outcome quotes

The total cost of an automation has five parts, and the model is usually the smallest:

• Model: cost per task at your volumes (the task table above), times the assumed 1.2x to 2x production overhead.
• Tools: fixed monthly fees for observability, vector storage and parsing, such as Langfuse Core at $29 a month (see the hidden-cost table).
• Hardware: only if you self-host, for example an owned DGX Spark at about ₹15,069 a month over an assumed 36 months (Worked example C).
• People: your own staff time to run and supervise the system; a quarter of one Indian AI engineer at the ₹10 lakh average is about ₹20,833 a month, salary only.
• Partner fees: if you use an implementation partner, the published bands in the previous section.

Partner quotes usually follow one of three methods, and many combine them. The table compares them from the buyer's side. The worked illustrations after it use hypothetical volumes and the benchmark prices above. They are illustrations of method, not quotes or a rate card.

Illustration 1: an outcome-priced support agent against running it yourself (hypothetical). Take 3,000 conversations a month on task T2 with Sonnet 5.5. If you run it yourself, the model and observability cost is 3,000 × $0.0522 × 2 (the top of the assumed production multiplier) + $29 of Langfuse observability = $342 a month, or $0.114 a conversation, before build and staff time. At the $0.99 Fin benchmark, the per-outcome bill depends on the share of conversations that resolve, which is an assumption:

• At a 50% resolution rate, the bill is 1,500 × $0.99 = $1,485 a month.
• At 30%, it falls to $891, but 2,100 conversations a month are left unresolved (arithmetic).
• At Salesforce's claimed 76% Fin average, it is $2,257.

On your own build, model choice moves the run cost: the same workload costs $45 a month on Haiku 5.5 and $655 on Opus 5.5 (arithmetic). Human escalation handling and support labour are not included in any of these figures.

Illustration 2: run costs of an invoice workflow (hypothetical). Take 1,000 three-page invoices a month for an Indian SME. Run costs are $0.75 of Haiku 5.5 tokens, $45 of Unstructured parsing once the 10,000 free pages are used, and $29 of observability. That totals $74.75, or ₹7,176 a month before build, support hours or GST (arithmetic). Your total adds the build and support time, whether from your own staff or a partner's fees. The run costs are small enough that people's time decides the total.

Partners often combine the methods in stages, and the sequence also suits a buyer:

1. Pay a fixed fee for the assessment and the pilot.
2. Measure your baseline during the pilot.
3. Move to a retainer plus pass-through usage, or to an outcome fee, once the baseline and resolution rate are known.

Two clauses decide who carries price changes. A price-review clause covers changes in model and tool prices, including the scheduled changes above. An exchange-rate clause applies wherever USD-priced inputs are billed to you in rupees. Read both before you sign.

The macro picture: capex up 80%, AI revenue catching up slowly

The vendors whose prices set your automation costs are spending heavily ahead of revenue. Their SEC filings show the scale.

Revenue is rising fast but from a smaller base. NVIDIA's quarter ended 26 July 2026 brought revenue of $96.2 billion (+106%) and data-centre revenue of $89.0 billion (+117%), with guidance of $108.0 billion for the next quarter (NVIDIA Q2 FY2027). Microsoft's AI business "surpassed an annual revenue run rate of $37 billion, up 123% year-over-year" in the March 2026 quarter (Microsoft FY26 Q3). Azure passed $100 billion for the fiscal year, and Microsoft 365 Copilot passed 30 million paid seats (Microsoft FY26 Q4). Microsoft's AI run-rate is about a third of its FY2026 capex (arithmetic: $37 billion ÷ $116 billion). Not every capex dollar is AI, and a run-rate is not annual revenue, so this compares disclosed numbers rather than forecasting anything. Salesforce reports Agentforce annual recurring revenue above $1.5 billion, though its definition widened in the same quarter (Salesforce Q2 FY27). Anthropic's latest company figure is run-rate revenue that "crossed $47 billion" (Series H announcement, 28 May 2026) (Anthropic Series H); about $65 billion at end-July is press-reported (a Bloomberg headline of 17 Aug 2026, article not read directly; Bloomberg via HN) and not confirmed by Anthropic. Financing is shifting toward capital markets. Alphabet raised $49.6 billion of equity and $20.3 billion of senior notes in the June quarter (Alphabet Q2 2026 release). Epoch AI states that "hyperscaler capex is on trend to outpace their cash inflows by the end of 2026" (Epoch AI data centres) and estimates a 1 GW AI data centre at about $38 billion of up-front capex (Epoch AI trends).

The bubble debate is live, and the sources here are mostly secondary. Wikipedia's summary reports Bank of England and IMF warnings of a "sharp correction" in AI valuations in October 2025. It also reports a May 2026 Financial Stability Board warning that a correction "could lead to sizeable credit losses to private credit investors". It cites a Morgan Stanley estimate of about $3 trillion of data-centre spending in 2025 to 2028, half from private credit, and OpenAI's roughly $1.4 trillion of data-centre commitments against about $13 billion of revenue (late 2025). On the other side, Goldman Sachs argued gains are "backed by actual profit growth", and Morgan Stanley called bubble fears "premature" (Wikipedia, AI bubble, secondary). On energy, the IEA estimate relayed by Wikipedia puts data-centre electricity at about 415 TWh in 2024 (about 1.5% of global use), rising to about 945 TWh by 2030 (Wikipedia, environmental impact of AI, secondary). Hardware inputs are also inflating. Conventional DRAM contract prices rose about 93% to 98% quarter on quarter in Q1 2026 (TrendForce data relayed by a secondary source) (IntuitionLabs). Apple raised Mac Studio memory prices, citing RAM costs (Notebookcheck), and the DGX Spark was reported at $4,699 in August 2026, up from a $3,999 launch price that rests on a search snippet (both secondary; see above).

For budgets, the implication is that today's API prices are set by vendors spending far more than they earn from AI. Some list prices are plausibly subsidised. The first signs of capex economics reaching customers are already visible: Gemini's 2027 rises, higher frontier prices at OpenAI and Google, and GPU rental prices rising rather than falling. A correction would most likely hit the financing layer first, because the debate centres on debt and private credit. Usage could keep growing even if valuations fell. Scenario detail is in Future outlook, and dates are in Timeline.

Risk notes: scheduled rises, ending promotions and the rupee

What this means for your business

Start with a baseline, not a model choice

Model tokens are the small part of the bill in text workflows. A support resolution costs fractions of a cent to about fourteen cents in tokens, while published outcome prices sit near a dollar or two. Most of your cost is integration, evaluation, escalation handling, monitoring and staff time. Measure a baseline, such as handling time per invoice, before you build; Gartner's forecast of cancelled agentic projects is a reason to ask any pilot for measured proof of value.

Decisions to make

• Routing. Build the prototype with the most capable model, set up evaluations, then move each step to the cheapest tier that holds accuracy (OpenAI, A practical guide to building agents). Routing simple turns to a Haiku- or Luna-class model can cut cost by 20x against a Sonnet-class model (arithmetic: $0.0522 ÷ $0.00267 for T2). Caching and batching each roughly halve the rest. Design around long-context thresholds such as Haiku 5.5's 100K boundary.
• API or self-hosting. At SME scale the API wins unless hardware runs near capacity. Self-host when data must stay in your environment, for example ahead of the DPDP obligations commencing on 13 May 2027, and budget hardware or rental, operations time and risk; it is not a saving. No vendor's own API processes requests in India; AWS Bedrock's India profiles cover the previous Claude generation, Kimi K3 and GPT-5.6 Terra/Luna, but not the newest models. If you qualify for IndiaAI compute, budget at the ₹153 list rate and treat the subsidy of up to 40% (subject to approval) as upside.
• Seats, outcomes or your own build. Compare outcome quotes with the $1 to $2 per resolved ticket band and seat quotes with the $20 to $30 cluster. Outcome pricing suits you only when the outcome is verifiable and defined in writing; expect minimums, such as Fin's 50 outcomes a month.

What to budget

Budget model cost at your volumes times the 1.2x to 2x overhead assumption, fixed tool fees that can exceed the model bill in a pilot, and staff time. Published US partner bands are $3,500 to $8,500 for an assessment, $18,000 to $60,000 for a build sprint and from $1,500 a month for managed service; for an Indian SME, a fixed-fee single-workflow pilot fits better (inference). Add 18% GST where quotes exclude it, and allow for the rupee, which has moved almost 6% this year.

Deadlines

• 21 November 2026: GPT-5.6 Sol's promotional price is guaranteed until then.
• 31 December 2026: WhatsApp accounts billed in India must move to INR billing; Microsoft's $18 Copilot Business promotion ends (secondary; dates conflict).
• 1 January 2027: Gemini 3.8, 3.7 and 3.6 Flash prices double.
• 13 May 2027: DPDP obligations commence.
• From 1 September 2027: earliest retirement dates for current Claude models.

What to test

In a pilot, log real token profiles, the resolution rate, long-context costs and, on DeepSeek, Indian traffic at peak rates. Keep exit plans for observability and orchestration tools, because prices and owners change.

Questions to ask any vendor or implementation partner

• Which model, tool and parsing costs pass through, at what rate, and can you see the vendor's list price for resold seats or credits?
• How do price-review and exchange-rate clauses work, and does every INR quote state GST?
• What counts as an outcome, what is the minimum, and what happens when a conversation passes to your staff?
• If the partner hosts a tool such as n8n for you, who holds the licence, and do its terms allow that use?
• Where is your data processed, and which models does that rule out?
• What is the exit plan if a tool changes owner or is archived, as Helicone, Langfuse and Flowise did in 2026?

Open questions and unverified items

Sources

1. X-Rates monthly average USD to INR, 2026
2. OpenRouter models API
3. Claude Platform pricing
4. Anthropic, Claude Opus 5.5
5. Anthropic, Claude Haiku 5.5
6. OpenAI API pricing
7. OpenAI API pricing (markdown)
8. OpenAI GPT-6 Astra model page
9. Google Gemini API pricing
10. Google Cloud Agent Platform generative AI pricing
11. xAI models and pricing
12. DeepSeek API pricing
13. OpenRouter rankings
14. NIST CAISI evaluation of DeepSeek V4-Pro
15. Artificial Analysis, gpt-oss-120b providers
16. a16z, Welcome to LLMflation
17. Epoch AI, LLM inference price trends
18. Epoch AI trends
19. Crypto Briefing, H100 rental costs
20. Claude models overview
21. DeepSeek Anthropic API guide
22. claude.com pricing
23. OpenRouter pricing
24. Cloudflare AI Gateway pricing
25. LiteLLM pricing
26. Langfuse pricing
27. LangSmith pricing
28. Helicone pricing
29. Braintrust pricing
30. Pinecone pricing
31. Chroma pricing
32. Voyage AI pricing
33. Unstructured pricing
34. Reducto pricing
35. LlamaParse pricing
36. n8n Sustainable Use License
37. Retell AI pricing
38. Bland AI pricing
39. LiveKit pricing
40. Vapi pricing
41. Plivo pricing
42. Sarvam AI API pricing
43. Synthflow pricing
44. Intercom pricing
45. Salesforce Agentforce pricing
46. Meta WhatsApp Business Platform pricing
47. TelecomTalk, WhatsApp Business price change
48. Authkey, WhatsApp marketing rate
49. Lambda pricing
50. Azure Retail Prices API, ND96isr H100 v5
51. Google Cloud VM pricing
52. IntuitionLabs, H100 rental prices
53. E2E Networks pricing
54. E2E Networks, H100 pricing in India
55. ECorpIT, India GPU cloud rental pricing 2026
56. Storyboard18, IndiaAI compute at ₹65 per hour
57. llama.cpp DGX Spark benchmarks
58. ainauten, DGX Spark price rise
59. Variety Infotech, NVIDIA listings
60. Tredence, AI engineer salary in 2026
61. Tech Insider, RTX PRO 6000 Blackwell price 2026
62. DPDP commencement notification, G.S.R. 843(E)
63. Microsoft 365 Copilot pricing
64. Microsoft Copilot Studio
65. Fin pricing
66. Zendesk pricing
67. Cursor pricing
68. Business Today, ChatGPT Go free for one year in India
69. Zapier pricing
70. n8n pricing
71. Botpress pricing
72. Dify pricing
73. Anthropic, Building effective agents
74. Salesforce signs definitive agreement to acquire Fin
75. Salesforce completes acquisition of Fin
76. Main & Machine pricing
77. Softomate, AI automation agencies London
78. Carmatec
79. Infyom, n8n workflow automation services
80. Antino, artificial intelligence
81. OptiWork
82. Main & Machine, AI implementation
83. US Bureau of Labor Statistics, software developers
84. Storyboard18, AI and ML engineer pay in India
85. Globant, AI Pods model
86. Gartner, over 40% of agentic AI projects will be canceled by end of 2027
87. SEC XBRL, Microsoft capex
88. SEC XBRL, Alphabet capex
89. Alphabet 10-Q, quarter ended 30 June 2026
90. SEC XBRL, Meta capex
91. SEC XBRL, Amazon capex
92. NVIDIA Q2 fiscal 2027 results
93. Microsoft FY26 Q3 press release
94. Microsoft FY26 Q4 press release
95. Salesforce Q2 FY27 results
96. Alphabet Q2 2026 earnings release
97. Epoch AI, data on AI data centers
98. Wikipedia, AI bubble
99. Wikipedia, Environmental impact of artificial intelligence
100. IntuitionLabs, HBM and DRAM AI memory demand
101. Notebookcheck, Mac Studio RAM pricing
102. X-Rates calculator USD to INR
103. Helicone is joining Mintlify
104. Flowise, The Future of Flowise
105. OpenAI, A practical guide to building agents
106. xAI API pricing
107. xAI regional endpoints
108. OpenAI, Data controls in the OpenAI platform
109. Moonshot, Kimi API pricing
110. Z.ai pricing
111. MiniMax pay-as-you-go pricing
112. Alibaba Cloud Model Studio pricing
113. NIST, CAISSI page (CAISI)
114. Claude extended thinking
115. OpenAI reasoning guide
116. xAI reasoning
117. Claude service tiers
118. AWS Bedrock model and Region compatibility
119. Google Document AI pricing
120. AWS Textract pricing
121. Plivo India voice pricing
122. Bolna pricing
123. TRAI-hosted comments on the draft TCCCPR Third Amendment (Exotel)
124. ChatMaxima, WhatsApp API pricing in India
125. Gallabox, WhatsApp pricing overview
126. MLPerf Inference v6.1 results (summary.csv)
127. AWS Price List, EC2 ap-south-1 (Mumbai)
128. E2E Networks, NVIDIA H100 product page
129. IndiaAI compute portal, price calculator
130. IndiaAI compute portal, eligibility criteria
131. Let's Data Science, RTX PRO 6000 price rise
132. GPU Poet, RTX PRO 6000 prices, July 2026
133. Techstory, Claude rupee billing in India
134. Office Watch, Microsoft 365 Copilot Business
135. GeoToolbox, Copilot pricing
136. eesel, ChatGPT Business pricing
137. fin.ai
138. Anthropic, Series H announcement
139. Hacker News listing of the Bloomberg report on Anthropic's run-rate (headline only)
140. Exotel blog: migrating from Twilio (search summary; not read directly)