Cost per task, not price per token: what GPT-6.2, Gemini 4 and Claude really cost
Two models with the same list price, $8 per million input tokens and $40 per million output tokens, cost $26.54 and $4.00 per task on the Artificial Analysis benchmark. Claude Fable 5.2 generates 275 million output tokens for it, GPT-6.2 Astra 65.8 million. The price per token does not show this difference.
Cost per task is the sum of all tokens a model consumes until it reaches a finished solution (input, cache, reasoning steps, answer), multiplied by the respective prices and divided by the number of tasks. Artificial Analysis calculates it this way for its Intelligence Index v5.0.1, a weighted average over nine evaluations. Cost per solved task additionally divides by the success rate. The token price is therefore one of three factors; consumption and success rate are the other two.
Price lists: what the vendors charge
All prices in US dollars per million tokens. Measurement carrier: the vendors' pricing pages, retrieved on 22 November 2026. These are vendor statements.
| Model | Input / output | Cache read | Source |
|---|---|---|---|
| Claude Fable 5.2 | 8 / 40 | 0.20 | Anthropic |
| Claude Opus 5.5 | 4 / 20 | 0.20 | Anthropic |
| Claude Sonnet 5.5 | 2 / 10 | 0.10 | Anthropic |
| GPT-6.2 Astra | 8 / 40 | 0.40 | OpenAI |
| Gemini 4 Neon | 3 / 15 | 0.15 | Google Cloud, Vertex AI, global |
| Gemini 4 Hydrogen | 1.50 / 7.50 | 0.075 | Google Cloud |
| Gemini 4 Carbon, Argon | 2 / 10 until 30 Nov 2026, from 1 Dec 2026 4 / 20 | 0.10, then 0.20 | Google Cloud |
At Google the lower price is a promotion: the price page describes it as a 50 percent credit on net spend, valid only until 30 November 2026. Anyone planning Carbon or Argon for December should calculate with $4 and $20. OpenAI cuts Astra by 20 percent against GPT-6 Astra and the cache price by 60 percent, according to its own page. Fable 5.2 and Astra then sit on the same price tier.
Measured independently: cost per task at Artificial Analysis
Artificial Analysis runs all evaluations itself. Cost per task is built from input, cache, reasoning and answer tokens times price, divided by the number of tasks and weighted by each evaluation's share in the index. The values come from the model pages, retrieved on 22 November 2026, index v5.0.1. The effort level in brackets is the one Artificial Analysis lists per model; it is not the same in every row.
| Model (effort level) | Index | Cost per task | Output tokens on the index |
|---|---|---|---|
| Claude Fable 5.2 (Max) | 65 | $26.54 | 275M |
| Claude Opus 5.5 (Max) | 59 | $10.04 | 218M |
| Claude Sonnet 5.5 (Max) | 55 | $10.02 | 335M |
| GPT-6.2 Astra (Max) | 62 | $4.00 | 65.8M |
| Gemini 4 Carbon (High) | 60 | $7.76 | 205M |
| Gemini 4 Neon (Xhigh) | 61 | $6.44 | 283M |
| Gemini 4 Hydrogen (Xhigh) | 54 | $3.16 | 304M |
| Gemini 4 Argon (High) | 50 | $9.95 | 113M |
Three calculations from this table, derived from the values listed:
- Fable 5.2 versus Astra: same list price, 26.54 divided by 4.00 gives a cost factor of 6.6 and a factor of 4.2 in output tokens (275 divided by 65.8). The index differs by 3 points.
- Hydrogen versus Astra: Hydrogen is 5.3 times cheaper per token (8 divided by 1.50) but only 1.27 times cheaper per task (4.00 divided by 3.16), with 8 index points less.
- Sonnet 5.5 versus Opus 5.5: half the token price, yet almost identical cost per task (10.02 versus 10.04), because Sonnet needs 335 million output tokens instead of 218 million.
The numbers describe this benchmark and say nothing about your tasks. For Carbon and Argon, Artificial Analysis uses the price tier of $4 and $20, so the promotion until 30 November is not in these figures. The effort level changes the calculation a lot. In a chart on the OpenAI page that shows Artificial Analysis values (vendor page, index v5.0.1, unchecked), Fable 5.2 costs $9.08 at High with an index of 61.8 and $26.54 at Max with 64.6. That is 2.8 points more for 2.9 times the cost.
Vendor calculations and what they leave out
In its Neon launch post of 18 November 2026, Google writes that Neon at High is 32 percent below Carbon's cost per task, and at xhigh more than 36 percent below competing models at their highest setting. These are vendor statements and we have not reproduced them. Artificial Analysis lists Neon at Xhigh and Carbon at High, so the 32 percent cannot be recomputed there.
OpenAI writes that Astra leads Fable 5.2 by 5.5 points on Agents' Last Exam "at just 5% of the cost". The chart data on the same page lets you check this: Astra at xhigh costs $2.89 with 78.3 percent, Fable 5.2 at Max costs $60.05 with 72.8 percent. That yields 4.8 percent of the cost and a 5.5-point lead. For Astra at Max the ratio is 7.4 percent. Both figures belong to the vendor page, and the text names no effort level.
Cost per task says nothing about reliability. The GPT-6.2 system card describes under Blocker Deception how often a model withholds a known obstacle in a coding task: 22.28 percent for GPT-6.2 Astra against 11.09 percent for GPT-6 Astra. OpenAI calls this a regression. With the Codex developer prompt the rate for Astra falls to 8.04 percent (chart on the launch page). According to the system card, that prompt does not apply to direct API requests. These are internal tests with deliberately placed obstacles, which OpenAI does not read as a production rate.
Stripe measures cost per run and a strict success rate
The Stripe evals rebuild integrations for payments, subscriptions and Connect. According to the methodology post, the average over all pairs is 78.1 percent of criteria, and the strict pass rate (all required criteria met) is 29.4 percent. Recomputed from the Stripe table, we get 29.4 for the strict rate and 78.2 for the score (simple mean over 14 pairs, with and without the Stripe plugin).
Measurement carrier: Stripe, vendor measurement, as of 19 November 2026, 69 scenarios in 2 runs, each with the Stripe plugin, default effort level, cost in dollars per run. The last column is derived by us: cost per run divided by the strict rate.
| Model (harness) | Score | Strict rate | Cost per run | Cost per strictly passed run |
|---|---|---|---|---|
| GPT-6.1 Sol (Codex) | 89.6% | 50.7% | $0.61 | $1.20 |
| GPT-6 Astra (Codex) | 90.0% | 52.9% | $3.63 | $6.86 |
| Claude Sonnet 5.5 (Claude Code) | 82.6% | 33.1% | $1.16 | $3.50 |
| Claude Opus 5.5 (Claude Code) | 86.2% | 36.0% | $3.15 | $8.75 |
| Claude Fable 5.2 (Claude Code) | 85.2% | 43.2% | $5.35 | $12.38 |
| Gemini 4 Argon (OpenCode) | 77.6% | 31.6% | $3.86 | $12.22 |
The caveats are Stripe's own. The graders are AI models (GPT-6.1 Sol for role play and grading; a cross-check with Opus showed no favoritism according to Stripe, Opus simply graded more leniently across the board), and the code is not executed: "A well-shaped but non-functional integration can score full marks". The harnesses differ per model, and Stripe warns against comparing across models. GPT-6 Astra was tested, version 6.2 is missing. The right-hand column assumes every failed run is worthless. That assumption sits behind every cost-per-solved-task figure, and it is rarely fully true. The spread from $1.20 to $12.38 stays visible all the same.
Without a meter there is no comparison
In a blog post of 28 August 2026, IDC reports that 67 percent of enterprises exceeded their agent budget by more than 10 percent in the past twelve months, and that only 45.4 percent have real-time dashboards for tokens and cost per workflow. The post does not name the survey wave for the 67 percent (Wave 4, July 2026, appears there next to a different figure). A second IDC post of 25 August 2026 says that the average token price fell by roughly 50 percent year over year in a fintech dataset, without bills falling. That quote comes from the Washington Post analysis and is no IDC measurement.
The tool changes the bill too. Google announced that it would shut down Gemini CLI for Pro and Ultra users on 18 June 2026 and replace it with the Antigravity CLI (agy); enterprise licenses stay unaffected. How far a harness moves the token bill for the same model is shown by the harness study.
What we measured ourselves
We did not measure cost per task ourselves. Everything above comes from Artificial Analysis, Stripe, OpenAI, Google, Anthropic and IDC. The token calculation of our gateway for Fable 5.2 is in Claude Fable 5.2: $8 per million tokens; it multiplies tokens by price and has no success rate. For the GPT-6.2 release, see the weekly edition IT News week 48; more posts are in the topic hub AI in production.
A measurement for your own team
Take twenty real tasks from daily work, two effort levels per model, and tests that actually execute the code. Record three numbers per run: tokens, cost, passed or not. From these, form the amount per passed run for each model and level, and assign costs to workflows before any model becomes the default. If you want to set up this measurement in-house, the framework is at Corporate LLM.
Further reading
- From the blog: Claude Fable 5.2: $8 per million tokens, access only with 30-day retention and Agent harness drives token costs up to threefold.
- Anthropic: Claude API pricing page, retrieved 22 November 2026. Vendor statement.
- OpenAI: Introducing GPT-6.2 Astra and Sol and GPT-6.2 System Card, retrieved 22 November 2026. Vendor statement, chart values read from the page source.
- Google Cloud: Vertex AI pricing (global region), Introducing Gemini 4 Neon and Transitioning Gemini CLI to Antigravity CLI, retrieved 22 November 2026. Vendor statement.
- Artificial Analysis: Intelligence Index v5.0.1 with the model pages, retrieved 22 November 2026. Independent measurement, live values can change.
- Stripe evals and methodology, as of 19 November 2026. Vendor measurement, AI graders, code not executed.
- IDC: The Agent Economy Is Scaling Faster Than It Can Be Metered and AI Tokenomics and the Cost Governance Gap, retrieved 22 November 2026. Analyst blog.
Why is the price per million tokens not enough to compare models?+
Because consumption per task varies by a large factor. According to Artificial Analysis, Claude Fable 5.2 (Max) generated 275 million output tokens on its index and GPT-6.2 Astra (Max) 65.8 million, at an identical list price of $8 and $40. The result is $26.54 versus $4.00 per task, a factor of 6.6. The index tasks differ from yours, and the mechanism applies to every bill.
What does cost per solved task mean?+
The cost of a run divided by the share of runs that actually deliver the result. Converting the Stripe eval numbers this way (derived, never published by Stripe in this form) gives $1.20 per strictly passed run for GPT-6.1 Sol, $6.86 for GPT-6 Astra and $12.38 for Claude Fable 5.2. The evals use AI graders and do not execute the code, so the values show an order of magnitude and no ranking.
How do I find the cost per task for my team?+
Take 20 real tasks from your own work, two effort levels per model, and tests that execute the code. For each run, record tokens, cost and pass or fail. In the end you have the amount per passed run for each model and level. A gateway log and a spreadsheet are enough to set this up in an afternoon.
senn-tech