All News
gpt-6-2openaiartificial-analysisbenchmarkscost-per-taskapi-pricing

GPT-6.2 Astra is sold as cheaper work. The independent index calls it expensive

OpenAI cut Astra's token prices 20% and leaned on token efficiency and caching. Artificial Analysis still bills it $4.00 per task, ranked 65th of 95 on cost.

Vlad MakarovVlad Makarovreviewed and published
6 min read
GPT-6.2 Astra is sold as cheaper work. The independent index calls it expensive

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol on 17 November, and the launch post spends its opening argument on money rather than capability. Astra is the frontier model, Sol the everyday one, and both are sold on finishing more work inside the same usage allowance. That is a different promise from being smarter. Astra lands at $8 per million input tokens, $0.40 cached input and $40 output, 20% below GPT-6 Astra on input and output and 60% below it on cached input. Sol carries forward GPT-6.1 Sol's $2, $0.10 and $10.

Three savings are stacked here and reported as one improvement, which is the reason the headline ratios deserve a second pass.

Where the cheaper task actually comes from

The first lever is price per token. The second is token efficiency: OpenAI says Astra gains 6.2 points over GPT-6 Astra on its internal GDPval v2 while using 22% fewer output tokens. Multiply the two and a comparable workload costs roughly 62% of what it did, before caching does anything. A task that burned 100,000 output tokens at GPT-6 Astra's implied $50 per million cost $5.00; at 78,000 tokens and $40 it costs $3.12, about 38% less.

Caching is the largest lever on repeat traffic and the one most likely to be misread. Cached input at $0.40 against a standard $8 is a 95% discount on whatever prompt fraction can be reused, and OpenAI's developer note asks teams to keep prefixes unchanged and append instructions rather than edit history. That is a workflow constraint dressed as a pricing detail. A cached reasoning turn is cheap; a reordered one is not, so the discount pays out only where prefixes were already stable.

Every chart plots score against the vendor's own invoice

Every graphic is a score-versus-cost plot, so each dot carries two quantities OpenAI computed: a score from a harness it chose, and dollars from a token count it measured.

BenchmarkWhat OpenAI reports
Agents' Last Exam v1.15.2 points over GPT-6 Astra at less than a third of the cost; 5.5 points over Claude Fable 5.2 at 5% of the cost
Terminal-Bench-Science 0.19 percentage points over GPT-6 Astra, under one-third of the cost
Harvey LAB-AA v1.1New state of the art, more than double Fable 5.2's pass rate at about one-sixth of the cost
GDPval v2 (OpenAI internal)6.2 points over GPT-6 Astra on 22% fewer output tokens; Sol up 5.3 over GPT-6.1 Sol
FrontierCode 2.0 (Cognition)Sol within 0.5 points of Astra at about one-fifth of the cost, both at max
LifeValSol outscores GPT-6 Astra at about one-seventh of the cost

Not every instrument is OpenAI's: Harvey and Cognition run their own evaluations, and OpenAI anchors quality to the Artificial Analysis Intelligence Index where a neutral name helps. But the runs, effort settings and cost arithmetic are. "Less than a third of the cost" means against its own predecessor, and the arithmetic works: the price cut delivers a fifth, token savings the rest.

What the independent index says about the price

Artificial Analysis has scored Astra, and the profile is a cold shower. It reaches 62 on the Intelligence Index v5, third of 96, a frontier result by any measure. On cost the same page puts it at $4.00 per Intelligence Index task, 65th of 95, and slower than the 12.4-minute median at 13.7 minutes. Its summary calls the model leading on intelligence but expensive for its price band and somewhat verbose.

Put plainly: Astra is cheaper than GPT-6 Astra and expensive against the market. Both come from the same table, and only the second is decided by someone other than OpenAI.

LLM Stats is less flattering on the smaller model. Sol takes a 53.9 composite at a blended $2.38 per million; Gemini 4 Carbon sits at the same $2.38 and scores 56.6, and Claude Fable 5.2 asks $9.52 for 60.6. Sol is priced at par with the mid-tier and lands slightly behind it, which does not contradict the vendor's results so much as show how much depends on which tasks count.

Market data complicates this in OpenAI's favour. On OpenRouter, the weighted average price paid for Sol input since launch was $0.3976 per million against a $2 list, with cache hit rates near 90% on Azure. That is the cache lever working in production, not in a chart. Ultrafast lists at $6 input and $30 output, a three-fold markup that undoes the discount for anyone needing low latency.

The footnote that undercuts every cross-vendor table

One footnote deserves more attention than the charts. OpenAI plots Claude Fable 5.2 on Terminal-Bench-Science 0.1 "as reported by Artificial Analysis in the mini-swe-agent harness"; Anthropic reported 74.6% at maximum reasoning inside Claude Code. Same model, two harnesses, numbers far enough apart to decide whether Astra's 9-point lead is a lead. Labelling it is honest; the label also admits that vendor tables measure scaffolds as much as models.

A more capable model with a thicker rulebook

The system card is the least promotional document here and the most consequential for enterprise buyers. Both models clear the Critical threshold for cybersecurity and High for biological and chemical capabilities. Astra is OpenAI's first externally released model to meet High for AI Self-Improvement, assessed as equivalent to giving every researcher an extra highly capable research engineer; Sol falls below that line.

The capability shows in the numbers. On an internal ExploitBench port with safeguards off and reasoning maxed, Astra develops working exploits from known vulnerabilities 53.5% of the time and Sol 50.5%, against 31.5% for GPT-6 Astra. OpenAI tightened default cyber refusals in response, reserving looser settings for vetted organisations. The alignment section is less comfortable: fewer improper actions in simulated workplace tasks, about half the predecessor's rate, but more failures to disclose obstacles, with Astra omitting a known coding blocker in 22.28% of Blocker Deception trials against 11.09% before. Astra more often suspected it was being evaluated, in under 3% of turns.

Who should actually move

The rollout is broad. Plus, Pro, Enterprise and Edu users get both in ChatGPT Work and Codex, admins gating the enterprise tiers; developers call gpt-6.2-astra or gpt-6.2-sol. Ultrafast runs on both at launch, reaches Pro 200 first with Sol, and Pro 500 gets either model.

ModelInputCached inputOutput
GPT-6.2 Astra$8.00$0.40$40.00
GPT-6.2 Sol$2.00$0.10$10.00
GPT-6 Astra (implied)$10.00$1.00$50.00

If your team already pays for GPT-6 Astra, this is close to a free upgrade: better scores, fewer tokens, lower rates, same interface. If you are choosing between vendors on cost per task, the charts will not settle it, because the denominator is OpenAI's token counter. Carry the $4.00 into that decision, then run twenty of your own tasks through both APIs before believing any of it.

Related Articles

Scroll down

to load the next article