GPT-6 Astra — OpenAI’s flagship model for hard end-to-end work. Photo: Unsplash.

OpenAI announced GPT-6 Astra on September 3, 2026 as its most capable broadly deployed model yet — built for difficult, multi-step work across reasoning, coding, computer use, browsing, research, and professional workflows. The API model id is gpt-6-astra. With a 1,050,000-token context window, up to 128,000 output tokens, and a knowledge cutoff of April 30, 2026, Astra is positioned as the successor flagship after GPT-5.6 Sol for teams that need longer horizons, stronger tool use, and fewer retries on the hardest tasks.

This guide covers verified specs, pricing (including long-context surcharges and Fast/Batch/Flex modes), reasoning effort levels, the Critical cybersecurity threshold under OpenAI’s Preparedness Framework, Codex’s experimental notes feature, who should upgrade from Sol, and how ChatGPT access differs from the API. Primary sources are linked throughout and listed in References.

Key specs at a glance

Astra targets frontier reasoning and agentic workloads. Photo: Unsplash.

OpenAI positions Astra for hard end-to-end work — not as a default for every chat. Official product pages emphasize efficiency (fewer tokens and retries) and a strong cost-efficiency frontier on professional and coding evaluations. On Terminal-Bench 4.0, OpenAI reports that GPT-6 Astra reached 57.9% versus 37.3% for GPT-5.6 Sol (OpenAI-reported figures; treat vendor benchmarks as claims to validate on your own workloads). See the launch post and work-focused overview.

Pricing deep dive (with worked examples)

Standard API rates for gpt-6-astra are published per 1 million tokens:

Cache writes are billed at 1.25× the uncached input rate ($12.50 vs $10). Once a reusable prefix (system prompt, tool schemas, large documents) is cached, subsequent hits bill at $1 per million — a 90% discount on that portion. Fast mode is a straight 2× price for up to roughly 2× speed. Batch and Flex are half of standard for workloads that can wait.

API pricing for Astra is driven by input, cache, and especially output tokens. Photo: Unsplash.

Worked example A — modest agent day (no long-context surcharge)

Assume 2,000,000 input tokens and 400,000 output tokens in one day, with 60% of input served from cache (1.2M cached / 0.8M uncached). Ignore cache-write setup for this steady-state day:

·       Uncached input: 0.8M × $10 = $8.00

·       Cached input: 1.2M × $1 = $1.20

·       Output: 0.4M × $50 = $20.00

·       Total ≈ $29.20 / day (~$876 / 30-day month)

Same volume with no caching: input $20 + output $20 = $40 / day. Caching here saves about 27% on the daily bill — and the savings grow as the reusable prefix grows.

Worked example B — one large research call

A single request with 100,000 input and 20,000 output tokens (under the 272K long-context threshold):

·       Input: 0.1M × $10 = $1.00

·       Output: 0.02M × $50 = $1.00

·       Total = $2.00 (standard)

·       Fast mode: $4.00; Batch/Flex: $1.00

Worked example C — long-context surcharge tripwire

Same shape as B but with 300,000 input tokens (above 272K) and 20,000 output. Long-context pricing applies to the full request at 2× input/cache and 1.5× output ($20 / $75 per million per reports):

·       Input: 0.3M × $20 = $6.00

·       Output: 0.02M × $75 = $1.50

·       Total = $7.50 (vs ~$4.00 if the same tokens were somehow billed at standard rates)

For ChatGPT seats, Pro/Business/Enterprise get GPT-6 Astra Pro (powered by GPT-6 Astra) with usage inside the subscription plus purchasable credits — fixed seat economics rather than pure per-token. For a secondary pricing walkthrough, see also Yotta Labs’ 2026 pricing post.

Context window and long-context surcharge analysis

Astra’s 1,050,000-token window is among the largest in OpenAI’s broadly deployed lineup, with a 128,000 max output. That is enough for multi-repo agent sessions, large document packs, and long tool traces — but the commercial design deliberately prices the “back half” of the window more aggressively.

According to OpenAI’s model docs, prompts with more than 272,000 input tokens are priced at 2× input and cache rates and 1.5× output for the full request. Practical implications:

·       272K is a billing cliff, not a hard capability cliff. Crossing it re-prices the entire call, including the first 272K tokens.

·       Prefer retrieval and selective attachment over dumping entire corpora when you can stay under 272K without hurting quality.

·       Prompt caching still matters below and above the threshold — but above 272K, cached rates are also doubled for that request.

·       Agent loops amplify cost. Resending a huge transcript every turn can flip many calls into long-context pricing; keep rolling summaries or Codex notes (below) in mind.

·       Max output (128K) is powerful and expensive at $50/M standard ($75/M long-context). Cap max_output_tokens for production paths that do not need novel-length answers.

Knowledge cutoff is April 30, 2026. For anything after that date, pair Astra with browsing / web search tools rather than relying on parametric memory alone.

Reasoning effort levels explained practically

Astra exposes reasoning.effort values: low, medium, high, xhigh, and max. Higher effort generally spends more internal reasoning tokens (and wall-clock time) before producing the final answer. A practical mapping for product teams:

OpenAI’s efficiency narrative for Astra is that better first-pass quality can reduce total tokens via fewer retries — so a higher effort setting can sometimes lower end-to-end cost even if a single call looks more expensive. Validate that claim on your eval harness; do not assume it for every prompt template.

Higher reasoning effort is most useful on multi-step coding and agentic tasks. Photo: Unsplash.

Safety and the Critical cybersecurity threshold

OpenAI’s safety overview states that GPT-6 Astra is its most capable broadly deployed model and the first to reach the Critical level of cybersecurity capability under the company’s Preparedness Framework. For users and admins, that designation is less about marketing and more about governance:

·       Capability ≠ permission. Critical cyber capability means OpenAI assessed high dual-use cyber potential; it does not mean Enterprise workspaces should enable the model casually.

·       Enterprise access is off by default at launch. Workspace admins must explicitly enable Astra — a deliberate control for security, legal, and risk teams.

·       Misalignment monitoring on tool-using inference is called out for Astra, reflecting the risk surface of agents that browse, shell, and operate computers.

·       Improved prompt-injection robustness versus GPT-5.6 Sol is an important claim for RAG and tool chains that ingest untrusted web or document content — still pair with allowlists, least privilege, and human confirmation on irreversible actions.

·       Rollout is staged: limited organizations first, then ChatGPT Plus / Pro / Business / Enterprise, API, Azure, and AWS Bedrock.

Administrators should update model allowlists, DLP policies, logging, and red-team prompts before flipping Astra on for production agents — especially computer-use or shell-enabled flows.

Astra’s Critical cyber rating elevates the need for admin controls and monitoring. Photo: Unsplash.

Codex and agentic workflows: experimental notes across context windows

One of Astra’s most operationally interesting launches is in Codex: a new way to preserve and retrieve context when the window fills. Historically, long coding sessions relied on compaction — summarizing prior work into a shorter blob. Each compaction can drop why a fix failed, exact error text, or how a component behaves.

With Astra, Codex can instead keep notes across context windows, preserving accumulated detail without repeatedly compressing everything into a single summary. Earlier context windows remain searchable, so Astra can recover requirements or test results from previous messages and tool outputs — even if those details were never written into its notes. OpenAI describes this as experimental at launch, with plans for it to become the default for Astra in the coming weeks.

Enablement is via Codex configuration (user-level ~/.codex/config.toml, or a trusted project’s .codex/config.toml). The documented experimental switch is along the lines of:

Key operational details to preserve (not just summarize):

·       Off by default at launch; start a new task after enabling so the feature attaches cleanly.

·       Designed to use notes + searchable history rather than repeated lossy compaction.

·       Eligibility and sign-in constraints apply (Codex docs reference ChatGPT sign-in on eligible consumer tiers for the experimental mode; check current Codex config reference before relying on Business/Enterprise/API-key paths).

·       Project-scoped config loads only for trusted projects — do not assume a repo-local override applies everywhere.

·       Notes are an aid, not a guarantee of perfect memory: verify critical facts (credentials, migration steps, failing tests) rather than trusting every persisted note blindly.

For teams building agents outside Codex, the product lesson still applies: invest in structured memory (notes, artifacts, searchable traces) instead of infinite transcript replay — especially with Astra’s 272K long-context surcharge.

Who should upgrade from GPT-5.6 Sol / prior models

Strong upgrade candidates

·       Engineering and SRE teams running long agent loops (migrations, multi-repo refactors, terminal-heavy workflows).

·       Products that need computer use, browsing, and research in one trajectory with fewer human interventions.

·       Workloads where OpenAI’s claimed Terminal-Bench 4.0 jump (57.9% vs Sol’s 37.3%, OpenAI-reported) would translate into fewer failed tool loops.

·       Orgs already hitting Sol’s quality ceiling on professional automation / document-heavy work and willing to pay flagship rates.

Consider staying on Sol (or routing mixed)

·       High-volume, low-stakes chat and classification where Sol’s lower list/promotional pricing wins on unit economics.

·       Latency-sensitive paths that do not need xhigh/max reasoning.

·       Enterprise tenants that have not completed security review — Astra may remain disabled until policy catches up.

·       Jobs that fit Batch on a cheaper model overnight rather than interactive Astra.

A sensible production pattern is multi-model routing: Sol/Terra/Luna (or open models) for the middle of the distribution; gpt-6-astra for the hard tail.

ChatGPT vs API access differences

Plus users may see Astra first inside specific surfaces (for example Work/Codex) depending on rollout phase; Pro/Business/Enterprise get the fuller GPT-6 Astra Pro experience subject to caps. Always confirm current entitlements in-product — staged launches change quickly.

FAQ

When was GPT-6 Astra announced?

OpenAI announced GPT-6 Astra on September 3, 2026.

What is the API model id?

Use gpt-6-astra (see the official model page).

How large is the context window?

1,050,000 tokens of context with up to 128,000 output tokens. Knowledge cutoff: April 30, 2026.

How much does GPT-6 Astra cost on the API?

Standard rates per 1M tokens: $10 input, $1 cached input, $12.50 cache writes, $50 output. Requests with more than 272K input tokens use 2× input/cache and 1.5× output for the full request. Fast mode is 2× price; Batch/Flex are 50% of standard.

What does “Critical cybersecurity capability” mean?

Under OpenAI’s Preparedness Framework, Astra is the first broadly deployed model assessed at the Critical cybersecurity capability level. Expect tighter defaults (including Enterprise off by default), monitoring on tool-using inference, and stronger admin scrutiny — not unrestricted cyber tooling for end users.

How do Codex notes work?

Experimental context management lets Astra keep notes across windows and search earlier messages/tool outputs instead of only compacting into summaries. Enable via features.context_management.experimental_mode in config.toml, then start a new task. Confirm current eligibility in Codex docs.

Should I replace Sol everywhere?

Usually no. Upgrade the hard, high-value trajectories; keep cheaper models for volume. OpenAI claims strong cost-efficiency on professional/coding evals and reports 57.9% on Terminal-Bench 4.0 versus Sol’s 37.3% — validate on your tasks before a blanket cutover.

Is Astra available on Azure and AWS?

OpenAI’s rollout path includes the API plus Azure and AWS Bedrock after the initial limited-org and ChatGPT phases. Check your cloud console for regional availability.

References

·       OpenAI — GPT-6 Astra: A new generation of intelligence

·       OpenAI — Safety overview: GPT-6 Astra

·       OpenAI Developers — GPT-6 Astra model docs

·       OpenAI — GPT-6 Astra: The next generation in intelligence for work

·       Yotta Labs — GPT-6 Astra pricing & API cost (2026) (secondary)

Rates and entitlements can change; verify against OpenAI’s live pricing and ChatGPT plan pages before budgeting.