OpenAI says GPT-6.2 Astra scored 9 percentage points higher than GPT-6 Astra at less than one-third the cost1. That is a cut of more than 67%. Artificial Analysis reportedly ran its own test and measured 36%.2 Vals AI measured 50%.3
Both are real savings. Both fall well short of the launch-day promise. On a $100,000 monthly model bill, OpenAI's claim puts the new cost under $33,333. The Artificial Analysis result puts it at $64,000.
That gap is about $30,667 a month, or roughly $368,000 a year. OpenAI introduced Astra and its cheaper sibling, GPT-6.2 Sol, on November 17, 2026,2 across ChatGPT Work, Codex and the API. We cannot tell you which number your workload will hit. Neither can OpenAI, and neither can any leaderboard.
The vendor claim haircut, measured
A vendor claim haircut is the discount you apply to a launch-day number until your own workload confirms it. Astra gives us three layers of figures, and sorting them shows where the haircut belongs.
| New bill on OpenAI's claimOpenAI · less than one-third the cost | $33,333 |
|---|---|
| New bill on independent resultArtificial Analysis · 36% cut | $64,000 |
| Monthly budgeting gapKoda analysis · claim vs test | $30,667 |
| Annual budgeting gapKoda analysis · twelve months | $368,000 |
Layer one is the token price. OpenAI says it cut input and output token prices by 20% and cached-input prices by 60%.4 Plotly lists Astra on OpenRouter at $8 per million input tokens and $40 per million output tokens, down from $10 and $50.5
Layer two is the task claim. OpenAI says Astra scored 5.2 points higher than GPT-6 Astra on Agents' Last Exam "at less than a third of the cost." According to OpenAI4, that benchmark tests long, multistep professional work.
Layer three is outside testing. Artificial Analysis measured $4.00 per task against $6.24,2 a 36% cut. Vals AI measured $10.68 against $21.28,3 about 50%. DataCamp reports Agents' Last Exam costs of $2.89 per task against $8.99,6 about a 68% cut, on the same benchmark OpenAI cited.
The haircut is the distance from layer two to layer three. The Artificial Analysis result lands 31 points below OpenAI's 67%. The Vals AI result lands 17 points below it. Plotly's data-analytics exam misses in the other direction: $4.11 per full run against $19.97,7 about 79%. One model, tested in one week, produced savings that more than double from the lowest test to the highest.
Why fewer tokens beat cheaper tokens
Think of the token price as a cab's rate per mile. Your bill is the fare, and the fare depends on the route. Astra charges 20% less per mile4, and on most of these tests it also drove a shorter route.
Artificial Analysis shows the split. On its standard 7:2:1 token mix, its blended token rate for Astra fell from $7.70 to $5.88 per million tokens8, about 24%. The task cost it measured fell 36%.2 Those extra 12 points came from Astra spending fewer tokens on each job.
BenchLeader reports output tokens dropping from 52,000 to 41,000,9 with reasoning tokens falling from 36,000 to 26,000. Plotly saw the same effect in time: Astra's median question took 12.4 seconds against 26.9 for GPT-6 Astra. Fewer tokens means a smaller bill, and any agent downstream waits less.
The route can also get longer. Microsoft Foundry lists long-context Astra at $16 per million input tokens and $60 per million output10, double the short-context input rate. Microsoft Foundry's EU data zone charges $48 per million output tokens for short-context deployments, 20% above the $40 headline. The DeepView comparison puts cached input on Anthropic's Claude Fable 5.2 at $0.20 per million4, half of Astra's short-context $0.40. (Token Tracker lists $8 input and $40 output per million tokens, scraped 14 November 2026.)
So the number you're really buying is cost per accepted task: total spend, retries and tool calls included, divided by the outputs you would ship without edits.
A model at 10 cents per attempt that passes 60% of the time costs about 17 cents per accepted task. A model at 14 cents with a 95% pass rate costs about 15 cents. The pricier call is the cheaper result.
Scale makes the haircut expensive to skip. At 10 million tasks a month on a $1.00 baseline, OpenAI's claim saves $6.7 million and the Artificial Analysis rate saves $3.6 million. That $3.1 million monthly difference is the price of believing the launch post.
The haircut takes three moves. Step one: treat OpenAI's 67% as the best case. Step two: budget on the Artificial Analysis figure of 36% until you have your own data2. Step three: let your measured cost per accepted task replace both numbers.
We think 36% is the right floor for a first budget, since it is the lowest major independent result. We can't say whether the gap between Artificial Analysis and Vals AI comes from the model or the test setup. Each evaluator picks its own prompts and its own pass bar. VentureBeat reported power users saying independent Codex testing found a flat result for Astra. The spread can run that wide.
Five more launch claims by 2027?
BenchLeader dates GPT-6 Astra to September 2, 2026,9 and lists GPT-6.2 Astra as added on November 17. That is 76 days between releases. If the pace holds, about five more Astra versions arrive before the end of 2027, since 409 days divided by 76 is about 5.4. Each one will ship with a launch-day cost claim.
Those claims compound in a budget. Two releases at OpenAI's claimed rate leave you paying about 11% of today's cost per task, since one-third of one-third is one-ninth. Two releases at the Artificial Analysis rate leave you at about 41%, since 64% of 64% is roughly 41%. A plan built on the first path undercounts spend almost fourfold against the second.
The risk is lopsided. Plotly's full Astra run cost $4.11.7 Mispricing a $100,000 monthly bill on the launch claim costs about $368,000 a year. A few dollars of testing protects a six-figure line item.
OpenAI's own results point somewhere bigger. According to OpenAI, GPT-6.2 Sol came within 0.5 points of Astra on FrontierCode 2.0 at roughly one-fifth of Astra's task cost2. For coding work, the comparison that matters may be Astra against Sol. Teams with a graded task set can route each job to the cheapest model that passes, then rerun that check on every release.
Our read: the graded task set is what compounds here. The model behind it may be replaced in 11 weeks. The set gets more useful with each swap, because every rerun adds one more dated price point you trust.
Price your top workflow per accepted task
The goal is one number: what a shippable result costs you on GPT-6 Astra, GPT-6.2 Astra and GPT-6.2 Sol.
Step one: pull 100 real tasks from last month's logs, all from your most expensive workflow. Real traffic carries your actual context lengths and cache hits, which no leaderboard sees.
Step two: write a pass rule for each task. A pass rule is a check that a script or a reviewer applies to say ship or reject. Keep it strict, because a loose bar makes every model look cheap.
Step three: run all 100 tasks through each model with the same prompts. Save the usage block from every API response, which breaks tokens out by type. Count every retry as spend.
Step four: price those tokens with your own deployment's rate card. A long-context job or an EU-zone job on Microsoft Foundry pays more than the OpenRouter rate Plotly cites. Then divide total spend by passed tasks.
Step five: compare your result with the haircut range. If your saving lands between 36% and 50%,3 budget at 36%.2 If it beats 50%, budget at 50% for a month and measure again in production. If it lands under 36%, look hard at Sol before you migrate anything.
The first run will break. A tool call will time out, or your grader will reject answers a human would accept. Fix the grader first, then rerun the whole set.
Plotly's exam cost about 10 cents a question on Astra7, so 100 analytics-style tasks would run roughly $10. Save the script so it reruns in one command. If the 76-day pace holds, the next Astra release lands around February 1, 2027, and your number will be ready before the launch post is.
Practise it step by step: Train the full skill in The Dojo
The bottom line
Budget on 36%2 and let your own grader set the real number.
Astra is cheaper on every independent test we found, and the size of that saving swings widely by workload and deployment. The launch claim is a best case. A graded set of 100 tasks costs about $10 to run and turns each future release into a measured price point. With the next Astra version possibly due around February 1, 2027, the script you save this weekend will produce your number before the next launch post appears.
