GPT-6 Astra: first on terminal engineering tasks, third tier on composite intelligence
gpt-6-astra
The OpenAI flagship tier, officially positioned for hardest end-to-end work (gpt-6-astra): 1.05M-token context, 128K-token single output, knowledge cutoff 2026-04-30, with Functions, Web search, File search and Computer use built in, so search, retrieval and interface control ship with the model instead of an outer agent framework. Reasoning effort runs low to max in five steps while price stays fixed at $10/$50, so the cost lever is token consumption; siblings Sol ($2/$10) and Luna ($0.1/$0.5) make a 100x spread, so load splitting can stay inside one vendor. Readings (2026-09-22): #1 on Terminal-Bench 4.0 at 58.18% (n_trials=330, pass@5 0.7121), the protocol closest to a coding agent's daily life; #2 on Arena Agent with net_improvement 11.54 and confirmed_success 17.70, both below leader Fable 5.1, but the field's highest praise at 32.79; #6 and #7 on the intelligence index (max 52.67). The cutoff is nearly five months older than this page, so version numbers and API changes must go through Web search; computer-use reliability appears on no board; closed, API only. Graded A (confirmed); not benchmarked by us.
- CONFIDENCE
- Confirmed
- Two or more independent sources, or reproduced by our harness
- KEY METRIC
- Terminal-Bench 准确率
- Confirmed · 2026-09
- MATURITY
- Product
- research → demo → product → production
Our takeWe grade it A (confirmed): OpenAI's official documentation supplies specs and pricing, and three boards synced into this site - Terminal-Bench, Arena Agent and the intelligence index - supply the readings. The two source groups are independent. Grade A is responsible for the factual layer only.
Its positioning is the clearest and the narrowest of the three at the top of our ladder: it is number one on the axis of getting work done hands-on. Terminal-Bench at 58.18% ranks first, and that board's protocol (a real terminal environment, n=330, scored on whether the task actually completes) is closer to a coding agent's daily life than a composite intelligence index; the built-in Web search, File search and Computer use surface then lets it run end to end without an agent framework bolted on. Those two facts together are the whole reason it sits third on our reasoning and code ladder.
Two places where it is not ahead have to be stated honestly. On the intelligence index its 52.67 is the lowest of the three, so anyone selecting by a composite board should not pick it. On the agent board it holds the highest praise score on the entire board (32.79) while losing net improvement to Fable 5.1 (11.54 against 13.71), and that split means its output more readily reads as "well done" while the actual forward progress is smaller. For production systems this is a bias worth actively guarding against: reviewers get carried away by how it looks, while the invoice and the diff only recognise net progress. One more: the knowledge cutoff of 2026-04-30 predates publication by nearly five months, so any task touching current versions must go through tools rather than memory.
The problem it solves: finishing a whole piece of work end to end, not answering one question well
OpenAI positions GPT-6 Astra as its flagship for the hardest end-to-end work. The weight in that phrase falls on "end-to-end": not producing a good reply to a good prompt, but taking ownership of an entire task - searching for context, reading files, calling tools, running code, inspecting results, deciding the next step - and continuing until the work is done.
The specs carry that positioning concretely. The model page lists the built-in tool surface directly: Functions, Web search, File search and Computer use, meaning "look things up, read a corpus, drive an interface" ship with the model instead of requiring an agent framework bolted on around it. Context is 1.05M tokens, single-response output is 128K tokens, the knowledge cutoff is 2026-04-30, and thinking effort is a five-position dial from low to max rather than an on/off switch.
Official specs and the sibling pricing ladder
| Model | API ID | Positioning | Price per million tokens |
|---|---|---|---|
| GPT-6 Astra | gpt-6-astra | Flagship, hardest end-to-end work | $10 in / $50 out |
| GPT-6 Sol | gpt-6-sol | Mid-tier workhorse | $2 / $10 |
| GPT-6 Luna | gpt-6-luna | Low cost, high concurrency | $0.1 / $0.5 |
Astra's own numbers: 1.05M token context, 128K token maximum output, knowledge cutoff 2026-04-30, five reasoning effort positions (low through max), and the built-in Functions / Web search / File search / Computer use tool surface. The spread across the three tiers is 100x from Luna to Astra, and the point of that ladder is load splitting inside a single vendor: high-frequency calls go to Luna, ordinary work to Sol, and only genuinely hard end-to-end tasks escalate to Astra.
Third-party water level: first on terminal engineering work, third band on composite intelligence
The three boards we sync (fetched 2026-09-22) give a picture that complements the Anthropic line almost exactly:
- Terminal-Bench 4.0: rank 1. Accuracy 58.18%, n_trials=330, pass@5 = 0.7121. Rank 2 is Fable 5.1 at 57.88% (pass@5 0.7879), rank 7 is another Astra row at 54.24% (pass@5 0.6667), rank 8 is Opus 5 at 53.94%. This board measures completing engineering tasks in a real terminal environment, making it the protocol closest to a coding agent's daily life of the three.
- Arena Agent board: rank 2. Net improvement 11.54, confirmed_success 17.70, praise 32.79. Both net improvement and confirmed success sit below leader Fable 5.1 (13.71 / 19.83), but praise is the highest on the board (32.79 against 31.83) - it wins "looked good" and loses "actually made progress", and that split has to be read honestly when selecting.
- Artificial Analysis intelligence index: ranks 6 and 7. Max effort 52.67, xhigh 52.39, below Opus 5.5's 57.62 / 55.99 / 53.58 and Fable 5.1's 53.35 / 53.20.
Taken together: Astra is number one on the axis of getting work done hands-on, and not number one on the axis of composite intelligence score. If your workload is coding agents, terminal automation, or tasks requiring the model to drive an interface itself, its ranking is the relevant one. If your workload is knowledge QA and long reasoning chains, Opus 5.5 at the same price scores 4.95 points higher on the index.
One data shape on Terminal-Bench needs clarifying: the same model appears on multiple rows (Astra at ranks 1, 2 and 7 with different readings) because that board records per submission and per configuration, not one row per model. We cite rank 1 at 58.18% and note n_trials=330 so readers know the sample-size convention.
The five-position dial: an engineering feature, not marketing
With reasoning effort spanning five positions and a fixed unit price ($10/$50 regardless of position), the cost lever sits on token consumption, not on unit price: same model, low effort burns fewer thinking tokens, max burns more, and the per-token price does not move. That is the opposite design philosophy from Anthropic's adaptive thinking - one hands the decision to the caller (predictable, tunable), the other to the model (convenient, unpredictable).
For production systems that is an advantage: you can route statically by task type (format conversion on low, architecture decisions on max) and draw your own cost curve instead of receiving one. The price of that is maintaining the routing table yourself and re-measuring it against real tasks periodically, because position semantics drift with model updates.
Boundaries
- Not advantaged on the intelligence index: 52.67 at max is last among the three entries at the top of our ladder. Anyone selecting by composite-intelligence board should not pick this one.
- High praise with lower net improvement is a warning sign: on the agent board it has the best subjective score on the whole board while losing net progress to Fable 5.1. In production, watch the latter; the former is easily over-read as a capability advantage.
- Knowledge cutoff 2026-04-30: nearly five months before publication, so anything involving latest version numbers or recent API changes must go through Web search rather than model memory.
- No board covers Computer use reliability: the built-in tool surface is a capability statement. We hold no interface-operation success-rate reading for it, so that part can only be understood as a vendor claim.
- Closed source, API only: no self-hosting, no fine-tuning, data leaves your perimeter.
Our verification status
Facts come from two independent sources: OpenAI's official model documentation (the models page on platform.openai.com, read directly for positioning, API ID, the five reasoning-effort positions, context and output limits, knowledge cutoff, built-in tool surface and three-tier pricing) and the raw rows of Terminal-Bench 4.0, Arena Agent and the Artificial Analysis intelligence index synced into this site (fetched 2026-09-22, including metrics field names and the n_trials convention). The two are independent, hence confidence A (confirmed).
What is not covered, stated plainly: we have not reproduced the capability claims. We have not measured Astra's Computer use success rate in our own harness, have not measured the cost-effectiveness curve of the five effort positions on our tasks, and have not verified whether Terminal-Bench's 58.18% reproduces in our terminal environment. The section on static routing by task type is a design suggestion drawn from the specs, not a measured conclusion. Reaching measured grade would require: an end-to-end completion comparison of Astra against Fable 5.1 and Opus 5.5 on one fixed task set, real token-spend and latency curves across the five positions, and a quantified Computer use success rate on interface-dependent tasks.