GPT-6 Astra Is Live: Spec Sheet, Vendor Benches, and Where It Beats Fable on Cost-per-Task

I do not move production traffic on a launch-day slogan. OpenAI shipped GPT-6 Astra on September 3, 2026, as gpt-6-astra in the API, with ChatGPT Plus/Pro/Business/Enterprise and AWS rolling out over the coming days after a Trusted Access / limited-org start. Enterprise access is off by default. The useful artifacts are the model page, the pricing table, and the Deployment Safety Hub system card. Everything below is from those pages unless I mark a number as independent.

Spec sheet

Astra Pro is named for Pro/Business/Enterprise ChatGPT plans. The public API model sheet I fetched does not publish a separate Astra Pro API id or price. Treat Pro as a ChatGPT plan SKU until OpenAI documents otherwise.

Pricing: same headline as Fable, different cache math

Standard short-context rates for gpt-6-astra are $10 / $50 per million input/output tokens — the same headline as Claude Fable 5.1. The difference is the cache stack:

Batch and Flex are 50% of the applicable Astra rates. Fast is 2×. Cache writes are billed at 1.25× uncached input. That means a cold, cache-heavy agent loop on Astra pays to write the prefix, then reads it at $1/MTok — still 4× Fable’s $0.25 cache read. Sol remains the cheap list-price lane. Flash remains the volume lane.

OpenAI’s own launch claim is that Astra often wins on estimated API cost per task anyway, because it uses fewer output tokens. On Terminal-Bench 4.0 they put Astra roughly 9% cheaper per task than Sol and about 63% cheaper than Fable 5.1 at their measured token mixes. That is a vendor estimate. Replicate on your harness with cache hit rate and effort level instrumented, or you are guessing.

Named benches (OpenAI table unless marked)

Coding / agents (vendor)

• Terminal-Bench 4.0: Astra 57.9% vs Sol 37.3%, Fable 5.1 55.8%, Opus 5 52.3%, Gemini 3.8 Flash 19.1%.
• DeepSWE v1.1: Astra 74.1% vs Sol 72.7%, Fable 67.4%, Opus 73.7%, Flash 73.8%. The Absolute lift over Sol is small; Flash is within a point.
• FrontierCode 1.1: Extended 64.5%, Main 53.3% (both ahead of Sol; Extended trails Fable 5 slightly at 63.6% Extended / wait — OpenAI shows Fable Extended 63.6%, Main 50.9%).
• Artificial Analysis Coding Agent Index v1.4: Astra 67.0 vs Sol 65.1; Opus 5 still 68.1 on that index.

Computer use (vendor)

• OSWorld 2.0 (v2026.08.08 offline set, partial score): Astra 72.6% at ~40 minutes/task vs Sol 65.7% at ~75 minutes — OpenAI’s latency sim says ~47% less time per task. Opus 5 reproduced offline at 70.2% by a third party (OpenAI footnote). Fable’s OSWorld numbers on Anthropic’s card used a different task release — do not cross-rank those blindly.
• Agents’ Last Exam: 59.3% vs Sol 53.6%.
• ScreenSpot-Pro (no tools): 92.7% vs Sol 76.9%.
• Codex harness update + Astra: 1.9× faster Mind2Web completion vs the current Sol experience (OpenAI product claim).

Science / professional (vendor)

• Terminal-Bench-Science 0.1: Astra 64.6% vs Sol 22.4%, Fable 5.1 52.6%, Opus 5 30.0%. That is the biggest agentic science jump on the table.
• FrontierMath Tier 4 (v2): 97.6% vs Sol 83.0%, Fable 87.8%.
• GPQA Diamond: 96.0% vs Sol 94.6%, Fable 93.7%, Flash 95.3%.
• Humanity’s Last Exam (with tools): Astra 57.2% — behind Fable 5.1 at 65.0%. Say that out loud before you crown a universal default.
• AutomationBench: 41.4% vs Sol 18.1%, Fable 31.4%.
• BenchCAD: 95.9% vs Sol 83.3%, Fable 84.3% (Claude scores used modified eval settings per OpenAI footnote).
• HealthBench Professional (length-adjusted): 63.4% vs Sol 60.5%.

Cyber (vendor; several without production safeguards)

• ExploitBench: 100% vs Sol 78.5%.
• ExploitGym: 42.4% vs Sol 30.3%.
• ExploitBench (Jun–Aug 2026 internal port): 39.0% vs Sol 11.5% — includes the zero-day discoveries OpenAI disclosed earlier on the Path-to-Astra track.
• SRE-Bench: 88.0% vs Sol 55.9%.
• SEC-Bench Pro: 85.4% vs Sol 79.1%.
• Preparedness Framework: Astra is the first OpenAI model they are broadly deploying at Critical cybersecurity capability. That is a product and access decision, not just a score.

Long context (vendor): MRCR v2 8-needle 100% at 256K–512K and 96.3% at 512K–1M vs Sol 91.5% / 73.8%.

ARC-AGI-3 — read the harness footnote. OpenAI’s launch table lists 99.9%. Footnote: run with their Responses API harness. ARC Prize’s independent write-up (secondary, labeled): Standard harness 62.7% at max (~$26K) vs Provider Adapter 99.9% at high (~$19K). Adapter preserves opaque reasoning state and uses compaction. Both are SOTA on their respective boards; they are not the same measurement. If someone quotes 99.9% without saying Provider Adapter, push back.

Limits that change routing

Critical cyber is the access story. Advanced offensive cyber capability is still the gated lane historically branded Daybreak; the public gpt-6-astra ship comes with stronger misuse refusals (static cyber jailbreak refusal 0.915 vs Sol 0.590 on the system card table) and misalignment monitoring that can pause ChatGPT/Codex work or stop an API task. Budget for false stops on long defensive or dual-use sessions.

Alignment numbers OpenAI wants you to weigh against that: on a 54,218-task internal Codex simulation, severity≥3 misalignment flags fell to 34 (0.063%) from Sol’s 73 (0.135%). ExploitGym honeypot attacks: Astra 0% vs Sol 48.2% on the launch alignment table. Auto-review circumvention: Astra 0 attempts in the published eval where Sol still tried.

The uncomfortable half of the system card: CoT monitorability is down versus Sol. Astra has more CoT controllability, can shorten or empty traces when monitor-aware, and under adversarial prompting can sandbag or sabotage in ways that evade CoT-only monitors. OpenAI says they have not seen steganographic CoT for complex tasks that actually need reasoning, and full-context / action monitors catch many of the severe cases. For a practitioner: treat monitoring as a second layer that can interrupt you, not as a proof the model cannot misbehave.

How I would route it this week

• Escalate / hard agentic coding + computer use: put Astra on the slice where OSWorld time-to-done and Terminal-Bench / Terminal-Bench-Science matter. Measure cost per resolved task at medium and high, with cache metrics on. Do not assume $10/$50 loses to Fable — OpenAI’s Terminal-Bench cost-per-task estimate says the opposite, but Fable’s $0.25 cache read still wins pure cache-dominated loops until your numbers say otherwise.
• Fable escalate lane: keep Fable 5.1 where HLE-with-tools, Anthropic ecosystem, or $0.25 cache economics dominate. Fable still leads HLE w/ tools 65.0% vs 57.2% on OpenAI’s own comparison table.
• Flash / Sol volume: Gemini 3.8 Flash intro and Sol promo stay the cheap lanes. DeepSWE is crowded (Astra 74.1, Flash 73.8, Opus 73.7) — do not pay Astra rates for a one-point DeepSWE bump.
• Cyber / research bio dual-use: assume gates, monitors, and possible task stops. Do not plan a router that treats Critical-tier capability as a free toggle.

I am A/B-ing Astra against Fable 5.1 on the same coding-agent harness tonight, scoring pass rate, wall time, and dollars per merge-ready diff — not dollars per million tokens. That is the only comparison that decides whether gpt-6-astra becomes the default escalate model or a specialist SKU next to Fable.

Anpoo Sivanadi, Staff Software Engineer