GPT-6.2 AstraGPT-6.2 SolLLM PricingPrompt CachingAI Benchmarks

GPT-6.2 Astra Cuts Its Price 20% and Its Cache 60%, Then Scores 77.1% on Terminal-Bench-Science for $7.15 a Task. Claude Fable 5.2 Scores 73.3% for $45.21.

OpenAI released two models on November 17. GPT-6.2 Astra is its top model, pitched as the best model for scientists, and it got cheaper: $8 / $40 per million tokens, 20% under GPT-6 Astra, with cached input down 60%. GPT-6.2 Sol keeps its $2 / $10 price and becomes the everyday model. On Terminal-Bench-Science, Sol now matches what GPT-6 Astra scored in September for about a twelfth of the cost per task.

2026-11-18·8 min read

The price sheet: the cache is the real cut

Per 1M tokensInputCached inputOutput
GPT-6.2 Astra$8.00$0.40$40.00
GPT-6 Astra (September 3)$10.00$1.00$50.00
Change−20%−60%−20%
GPT-6.2 Sol$2.00$0.10$10.00
GPT-6.1 Sol$2.00$0.10$10.00

OpenAI’s launch post; GPT-6 Astra’s prices from its September launch, which we covered in GPT-6 Astra’s price hike.

The headline says 20%. The line that matters for agents is the middle one. GPT-6 Astra discounted cached input by 90%; GPT-6.2 Astra discounts it by 95%, the same ratio Sol already had. Coding agents and long chats reread the same context on every turn, so most of their input tokens are cache hits. Our arithmetic for one long agent session with 10M input tokens, 92% of them cached, and 200K output tokens:

10M in (92% cached), 200K outUncachedCachedOutputTotal
GPT-6 Astra$8.00$9.20$10.00$27.20
GPT-6.2 Astra$6.40$3.68$8.00$18.08
GPT-6.2 Sol$1.60$0.92$2.00$4.52

Our math from OpenAI’s list prices; real sessions vary.

On that session Astra gets 34% cheaper, not 20%, because the cached tokens were the biggest line on the bill. That is also why heavy users on the $200 ChatGPT plan, which draws the same kind of sessions in Codex, have reason to watch how far Astra now stretches. OpenAI’s docs say a cached prefix stays reusable for at least 30 minutes and may be kept longer; on GPT-5.6 and later models you set the minimum lifetime with prompt_cache_options.ttl.

Terminal-Bench-Science: 77.1% for $7.15

OpenAI backs the “best model for scientists” claim with two science benchmarks. Terminal-Bench-Science 0.1 gives a model scientific research workflows to run in a terminal, and OpenAI plotted every model at each reasoning setting against its cost per task:

Line chart of Terminal-Bench-Science 0.1 score against cost per task: GPT-6.2 Astra climbs from 56% at about $2.50 to 77.1% at $7.15; GPT-6.2 Sol reaches 68.0% at about $2; GPT-6 Astra tops out at 68.1% at $23.80; GPT-6.1 Sol at 57.0% at about $5.40; Claude Fable 5.2 at 73.3% at $45.21; Claude Opus 5.5 at about 63% for about $23
Redrawn from OpenAI’s launch chart. Points are read off the chart; the top point of each line matches OpenAI’s table. OpenAI ran Fable 5.2 in the mini-swe-agent harness.
ModelTerminal-Bench-Science 0.1Cost per taskVals Science Index v1
GPT-6.2 Astra (max)77.1%$7.1568.4%
Claude Fable 5.273.3%$45.2163.7%
GPT-6 Astra68.1%$23.8065.1%
GPT-6.2 Sol68.0%about $258.5%
GPT-6.1 Sol57.0%about $5.4059.6%
Claude Opus 5.5about 63%about $23—

OpenAI’s tables as reported by DataCamp; “about” values are read off OpenAI’s chart.

Astra gains 9 points over GPT-6 Astra at 30% of the cost per task, and beats Fable 5.2 by 3.8 points at about a sixth of its cost. Sol lands on GPT-6 Astra’s score for roughly a twelfth of the money. One catch in the same table: on Vals Science Index v1, the new Sol (58.5%) scores a point below GPT-6.1 Sol (59.6%).

The whole GPT-6.2 line sits left of every other model on the chart. The cheapest Astra setting that beats Fable 5.2’s best score costs about $5 a task.

GDPval v2 and work benchmarks: Gemini 4 Carbon keeps up

OpenAI also introduced GDPval v2, the next version of its benchmark of economically valuable tasks. It’s internal, so the models are, if anything, tuned toward it; the gain over OpenAI’s own predecessors tells you more than the gap to rivals.

BenchmarkGPT-6.2 AstraGPT-6.2 SolGPT-6 AstraFable 5.2Gemini 4 Carbon
GDPval v2 (OpenAI, internal)50.4%46.9%44.2%44.8%48.8%
Agents’ Last Exam v1.178.3% ($2.89)73.0%73.1% ($8.99)72.8% ($60.05)77.4% ($4.60)
AutomationBench 1.0.652.4%41.2%41.4%51.4%53.9%
Harvey LAB-AA v1.115.0%11.1%8.6%7.2%—
FrontierCode 2.044.2% ($5.97)43.8% ($1.29)42.8%49.1%—
BenchCAD 2.1 Core82.5% ($5.43)73.1%62.6%83.5% ($50.77)—

OpenAI’s tables as reported by DataCamp; cost per task where given.

On GDPval v2, Astra gains 6.2 points over GPT-6 Astra while writing 22% fewer output tokens, and Sol gains 5.4 over GPT-6.1 Sol. Gemini 4 Carbon is the surprise in OpenAI’s own tables: it beats Fable 5.2 on GDPval v2 and tops AutomationBench outright, at $4.60 a task on Agents’ Last Exam to Fable’s $60.05. Fable keeps the top spot on FrontierCode and BenchCAD, at several times the cost.

What independent testers found

Artificial Analysis Intelligence Index v5.0.1ScoreCost per task
Claude Fable 5.264.6$26.54
GPT-6.2 Astra62.1$4.00
Gemini 4 Carbon60—
Claude Opus 5.559—
GPT-6.2 Sol57.3—
GPT-6 Astra56.9—

Artificial Analysis puts Astra second behind Fable 5.2, 5 points above GPT-6 Astra, and Sol level with GPT-6 Astra at about 15% of its cost per task. Two catches: on its Coding Agent Index Astra didn’t move and Sol dropped 3 points from GPT-6.1 Sol, and on AA-Omniscience at max effort Astra hallucinated on 59% of the questions it couldn’t answer and Sol on 55%. On its 3D benchmark AA-3D, Astra scores 1,318 Elo to Fable 5.2’s 1,271 and Sol’s 1,259.

Vals AI is cooler: Astra 7th and Sol 9th on the Vals Index, Fable 5.2 first. Its cost per task tells the price story again: Astra $10.68 to GPT-6 Astra’s $21.28, and Sol $3.18. Epoch AI puts Astra at 174 on its ECI, tied with Fable 5.2.

Cybersecurity: a jump, and stricter safeguards

On ExploitBench (OpenAI’s internal port), Astra scores 53.5% and Sol 50.5%, up from 31.5% for GPT-6 Astra and 21.5% for GPT-6.1 Sol. On SecureSWE Astra goes from 48.7% to 65.7%. Both models reach the Critical cybersecurity level of OpenAI’s Preparedness Framework, and per the system card Astra is the first released model to reach High for AI self-improvement. OpenAI ships tighter cyber safeguards with them and says they will also block some legitimate security research, so expect more refusals on exploit work.

Where you get it, and Ultrafast on the $200 plan

In ChatGPT both models are live for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex; Enterprise and Edu admins switch them on. In the API they are gpt-6.2-astra and gpt-6.2-sol, with Ultrafast mode available on both. The $500 Pro plan gets Ultrafast on both models, and the $200 Pro plan gets Ultrafast for the first time, starting with Sol.

What this means if you build with AI

Long agent sessions: re-run the math for Astra. If you dropped GPT-6 Astra over cost, the cached-input cut changes the bill more than the 20% headline: a cache-heavy session comes out about a third cheaper.

Everyday calls: GPT-6.2 Sol at the old price. It reaches GPT-6 Astra’s science score and Intelligence Index score at a fraction of the cost. For coding agents, test it against GPT-6.1 Sol first: Artificial Analysis measured a 3-point drop on coding.

Hardest code and CAD: Fable 5.2 still leads, but on Terminal-Bench-Science it costs about 6x Astra per task for a lower score. Pick by cost per task on your own workload, not by the price sheet.

Keep the model name in config. With releases every few weeks, a switch should be one line, not a code change.

On MegaBrain, GPT-6.2 Astra and Sol aren’t in the catalog yet. Their predecessors, GPT-6 Astra and GPT-6.1 Sol, are available through one OpenAI-compatible API with one balance, next to Claude and Gemini models and the image and video models we cover here. If you run a coding agent such as OpenCode on that key, the OpenCode setup guide shows the config.

Sources

Pages opened November 18, 2026. Every number above comes from these pages, or from our arithmetic on them where we say so.

Full disclosure: MegaBrain sells pay-as-you-go API access to image, video and text models, including GPT-6 Astra, GPT-6.1 Sol and Claude models. MegaBrain is not affiliated with OpenAI, Anthropic, Google, Artificial Analysis, Vals AI or Epoch AI.

MegaBrain for Mac · an AI coworker

Give MegaBrain the task, not the steps.

It works on your files, writes documents, spreadsheets and decks, connects to your team’s tools and asks before anything risky. Free to start.

Newsletter

Stay in the loop

Get our latest AI research and guides. No spam, unsubscribe anytime.