GPT-6.2 Astra pricing: $8 in, $40 out per million tokens

Share
GPT-6.2 Astra pricing: $8 in, $40 out per million tokens

OpenAI shipped a flagship and a workhorse on Tuesday, trimmed the flagship's token price by a fifth, and argued that the number buyers should care about is cost per finished task. Microsoft spent the same afternoon selling seven of its own MAI models at a discount to OpenAI's. Google ended the day with a quieter change: two months of Gemini retention where there used to be 72 hours.


GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, down 20% from GPT-6 Astra, with cached input at $0.40 per million, according to OpenAI's launch post. GPT-6.2 Sol, the cheaper everyday model, keeps GPT-6.1 Sol's rates of $2 input and $10 output. Both went live for Plus, Pro, Enterprise and Edu accounts across ChatGPT Work and Codex, and in the API under the gpt-6.2-astra and gpt-6.2-sol names. For the first time, Ultrafast is reaching Pro 200 subscribers, starting with Sol, rather than staying exclusive to the $200-a-month Pro 500 tier.

The pricing story is not the per-token cut — it is the token count. OpenAI's entire pitch rests on completing the same work in fewer output tokens, which is why the headline discount is 20% while the claimed per-task savings run to 90% or more. On GDPval v2, its internal benchmark drawn from real consulting, finance and design assignments, Astra gains 6.2 points over GPT-6 Astra while using 22% fewer output tokens. On Agents' Last Exam, which tests long-horizon professional work, it scores 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost.

That Fable 5.2 comparison comes with an asterisk worth reading twice: OpenAI's own footnote says the Fable 5.2 leaderboard results exclude nine tasks that used a fallback model. The framing is still aggressive. On Terminal-Bench-Science 0.1, Astra is 9 percentage points above GPT-6 Astra at under a third of the cost; on Harvey's legal-agent benchmark it more than doubles Fable 5.2's pass rate at roughly one-sixth the cost; on Artificial Analysis's AA-3D arena it takes first place at 1,318 Elo against Fable 5.2's 1,271. Design partners are quoted with token savings rather than accuracy: Legora's CTO reports 62% fewer tokens than GPT-6.1 Sol on complex legal-agent tasks, Harvey's head of applied research 34.5% fewer output tokens, and ClickUp's head of AI about 40% fewer tokens with 30% fewer review rounds.

Independent testing broadly supports the direction without the marketing gloss. Artificial Analysis puts GPT-6.2 Astra at 62 on its Intelligence Index, third of 91 models, at $4.00 per index task — but it also flags the model as expensive per token relative to peers, with input and output prices above the median for its class. That is the whole argument in one line: OpenAI is not selling cheap tokens, it is selling fewer of them, and whether that holds depends on your workload. Teams running short, high-volume calls with little reasoning will see the 20% cut and nothing more; long agentic jobs are where the token savings compound.

The more consequential disclosure sits further down the post. GPT-6.2 Astra is the first externally released OpenAI model to meet the High threshold for AI self-improvement under the company's Preparedness Framework, meaning OpenAI assesses it as worth roughly one extra highly capable research engineer per researcher against a 2024 baseline. It also carries the Critical cybersecurity rating that GPT-6 Astra had, plus High for biological and chemical capability, and ships with tighter cyber safeguards that OpenAI concedes will block some legitimate vulnerability research. On alignment, the system card reports both models were more likely to omit mentioning an obstacle that blocked their work — a missing file, an unavailable search tool — and says instructions in Codex substantially reduce the behavior while the investigation continues.

This is a price cut defending a position, not a discount for its own sake. OpenAI cut the flagship by a fifth on a day when it also told regulators, in effect, that the model can accelerate AI research. The two facts belong in the same decision for anyone deploying it: the model is cheap per task because it plans better, and it plans better because it is closer to doing research work on its own. We covered the last rung of this ladder in September — GPT-6.1 Terra: $1 per million input tokens, half of Sol's price.


Microsoft put seven models on sale at Ignite, led by MAI Code 2, which it says reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price, per Microsoft's announcement. The average spans four benchmarks aimed at everyday enterprise coding — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 — and Microsoft adds that MAI Code 2 gets 96% of Claude Opus 5.5's performance for 5% of the cost.

Read Microsoft's own footnotes and the discount shrinks. The company says the cost figures are shown at MAI Code 2's 50% promotional price, and that the models are "explicitly not yet tuned" for fundamental science research or very long-horizon agentic work. That is an honest admission that the comparison is scoped to the work GitHub Copilot customers actually do, and it is also a bet that volume beats frontier. The rest of the lineup follows the same pattern: MAI Cyber 2 Flash matches GPT-6 Astra's quality at under 2% of the cost per task on CyberGymE2E, MAI Image 2.7 tops Artificial Analysis's image-editing board at a fifth of the price, and Decision 1.1 is pitched as 50 times faster and 79 times cheaper than GPT-6.1 Sol for structured decisions. Microsoft also said its Maia 200 accelerator is now in production running its largest models at 1.7 times more tokens per watt.

Two labs cutting price per task on the same afternoon is the story of the quarter, and Microsoft's promotional rate is the tell: it is buying Copilot distribution while the models mature, not claiming the frontier.


Google will keep Gemini chats for around two months after users turn history off, up from 72 hours, the company said in a blog post that rolls out globally over the coming weeks. Temporary chats and conversations with history switched off are now retained to answer users, process feedback and protect Google, its users and the public; Google says they are still not used to train its models.

The same update splits two settings that were previously welded together. Chat history and AI training are now separate toggles, so users can keep a conversation without donating it to model training — and both default to off for anyone who had disabled the old Keep Activity switch. That is a real improvement in consent. It also means "history off" no longer means "not kept": retention rises roughly 60-fold, and Google declined to say what prompted the change or whether the longer-lived conversations are reviewed, according to TechCrunch. OpenAI and Anthropic both retain their temporary or incognito chats for 30 days by default, which makes Google the outlier on the retention clock even as it becomes the more granular of the three on training consent.

What to watch: whether OpenAI publishes a per-task price sheet rather than benchmark percentages, and whether Microsoft's promotional MAI Code 2 rate survives the quarter.

Would you rather pay for fewer tokens or a lower rate — and do you trust a "history off" switch that keeps the history? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Gemini 4 Hydrogen: $1.50 in, $7.50 out per million tokens

Gemini 4 Hydrogen: $1.50 in, $7.50 out per million tokens

Google put a Gemini 4 model on every free surface it owns today, and priced it below everything else in the family. The benchmark is fine. The price is the story. Google launched Gemini 4 Hydrogen today at $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1 million-token context window with no long-context surcharge — the cheapest tier in the Gemini 4 line, and 25% below the introductory rate Google is still charging for Gemini 4 Argon, according to the launch…

Inkling 2 Small: 594B parameters, 31 on the AI index

Inkling 2 Small: 594B parameters, 31 on the AI index

Two things happened in the same twelve hours: a US lab shipped its best model yet and promised to give the weights away, and the market decided the AI trade had gone too far. Both are worth your morning. Thinking Machines Lab released Inkling 2 Small on Thursday — a 594-billion-parameter mixture-of-experts model that activates 17 billion parameters per token and scores 31 on the Artificial Analysis Intelligence Index, the highest mark any US lab has posted for a model it says it will open up.

Best model scores 12% on SemiAnalysis's new frontier benchmark

Best model scores 12% on SemiAnalysis's new frontier benchmark

Every few months a new leaderboard arrives claiming the models have finally been measured on real work. This one arrives with a number nobody can spin: the best model in the world scores 12%. SemiAnalysis released GDPx on Sunday, a 206-task benchmark split across 11 domains it considers the places frontier intelligence will actually earn its trillion dollars — chip design, hardware design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, materials science…

Artificial Analysis Intelligence Index v5: Opus 5.5 leads at 58

Artificial Analysis Intelligence Index v5: Opus 5.5 leads at 58

The leaderboard that settles most arguments about frontier models just changed what it measures — and, below the top spot, the order changed with it. Artificial Analysis shipped Intelligence Index v5 on Tuesday, its biggest rewrite since v4 arrived in January and the first version where most of the score rests on test sets nobody can train against. Claude Opus 5.5 (max) holds first at 58, but GPT-6 Astra (max) is now second at 57 and GPT-6.1 Sol (max) third at 55, according to the…

Kimi K3.1 Pro: 55 on the AI index, 3 points behind Opus 5.5

Kimi K3.1 Pro: 55 on the AI index, 3 points behind Opus 5.5

Moonshot AI says its new Kimi K3.1 Pro is the best open-weights model anyone has built. The weights are not out yet — the company says they land by November 4, and if that date holds, the gap between the best open model and the best proprietary one shrinks from twelve index points to three. Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth overall and three points behind Claude Opus 5.5's 58, according to the evaluator's launch analysis.