GPT-6 Astra release timeline separating OpenAI rollout facts from EvoLink route verification

model-release

GPT-6 Astra Release Date: September 3, Specs & Rollout

EvoLink Team

Product Team

July 27, 2026

Updated on September 5, 2026

17 min read

Update: GPT-6 Astra is callable on EvoLink as gpt-6-astra at 10% below OpenAI list price, on the same API key you already use. Pricing, model ID, and code samples are on the GPT-6 Astra API page; integration details are in the GPT-6 Astra API guide.

GPT-6 Astra was released on September 3, 2026. OpenAI documents the model ID gpt-6-astra, a 1.05M-token context window, 128K maximum output, text and image input, and $10/$50 list pricing per million input/output tokens. Access began with a limited set of organizations on September 3 and broadened to the OpenAI API and paid ChatGPT plans from September 4.

This article owns the release date, the rollout timeline by channel, and the record of which pre-launch claims were right or wrong. It is not the pricing or integration page.

"Release Date" Can Mean Three Different Events

Search results often collapse three milestones into one date. They should be checked separately:

Milestone	What would count as confirmation	What would not count
Product announcement	Confirmed September 3 by OpenAI's GPT-6 Astra release and safety posts	Earlier leaks, prediction markets, screenshots, and unnamed systems
ChatGPT availability	Rolling out to Plus, Pro, Business, and Enterprise over the following days	A release headline without account-level access
API availability	OpenAI documents `gpt-6-astra`, API surfaces, pricing, and limits; account access still follows rollout controls	An OpenAI listing does not prove access on every cloud or gateway
Gateway availability	Each channel publishes its own listing date and price: Azure Foundry and OpenRouter on days 0–1, EvoLink after route, pricing, and billing checks	A provider announcement naming a cloud (for example Amazon Bedrock) before that cloud lists the model

A model can be released before every account receives access, and an upstream API can be live before every cloud and gateway lists it. Those are separate states, not contradictory reports.

The Direct Answer: What Is Confirmed Today?

Question	Answer as of September 5, 2026	Evidence level
Is GPT-6 released?	Yes. OpenAI released GPT-6 Astra on September 3, 2026	Official release
What is the upstream model ID?	`gpt-6-astra`	Official API documentation
Was it broadly available on day one?	No. September 3 covered a limited set of organizations; API and paid-plan access broadened from September 4	Official rollout statement
What are the context and output limits?	1.05M context and 128K maximum output tokens	Official model documentation
What is OpenAI list pricing?	$10 input, $1 cached input, $12.50 cache write, and $50 output per 1M tokens; 2x input and 1.5x output above 272K input tokens	Official pricing page
Is it available on EvoLink?	Yes. Call `gpt-6-astra` with your existing EvoLink API key at 10% below OpenAI list price	EvoLink route verification
Is it on Azure and Amazon Bedrock?	Azure Foundry: generally available since September 3. Amazon Bedrock: not listed as of September 5, although OpenAI named it as a channel	Azure and AWS documentation

The release date is settled. The remaining questions concern rollout eligibility per account and which clouds and gateways have actually listed the model.

What the GPT-6 Astra Launch Confirms—and What It Does Not

The launch resolves the entity, core specification, and upstream commercial contract. It does not collapse every provider into one availability state:

OpenAI now confirms	The launch does not confirm
The product name is GPT-6 Astra and the model ID is `gpt-6-astra`; there is no generic `gpt-6` alias	Access for every OpenAI account tier or region on day one
1.05M context (922K maximum input), 128K output, text/image input, and text output	Audio/video support or a memory API
$10 input, $1 cached input, $12.50 cache write, and $50 output per 1M tokens; Batch and Flex at 50%, Fast at 2x	Third-party gateway prices, which each channel publishes separately
Responses, Chat Completions, Batch, streaming, and structured outputs; tool calling on the Responses API	Tool calling on Chat Completions, which is not supported (OpenAI documents Responses only)

OpenAI's evaluations are meaningful vendor evidence for difficult coding, computer use, research, and agent work. They are not a universal production verdict; teams still need repeatable workload tests and route-specific cost data.

GPT-6 evidence workflow filtering official sources, quarantining unverified claims, and keeping the production route locked until verification

How to Judge a GPT-6 Claim

The evidence ladder prevents a familiar failure mode: repeating a precise-looking claim until repetition is mistaken for confirmation.

Evidence tier	Examples	What it can establish
Tier 1: provider product and API records	OpenAI model catalog, product posts, pricing, API docs, release notes	Name, status, model ID, supported surface, commercial terms
Tier 2: provider executive or research statements	Named OpenAI speaker, research result, security disclosure	Direction or existence of an unnamed system, limited to the exact wording
Tier 3: attributed reporting	Reputable publication with named sources or direct observation	A reported event or private plan, still subject to change
Tier 4: markets and community discussion	Polymarket, Reddit, X, forums	User expectations, questions, and sentiment—not provider facts
Tier 5: unattributed SEO pages and screenshots	Pages with no primary link, cropped interfaces, copied “spec tables”	A lead to investigate; nothing should be treated as confirmed

Historical absence still needs careful wording. “No GPT-6 entry appeared in the pages reviewed on August 5” describes the old snapshot; it must not be repeated as the current state after the September 3 release.

Verified OpenAI Signals, in Order

These are the events that actually happened, each with a source you can check:

August 2025: Altman discusses the direction of the next model

In remarks reported by CNBC, Sam Altman said the gap to the next flagship would be shorter than GPT-4 to GPT-5, and that memory and personalization are the core demand: "People want memory." This is one of the few executive remarks to use the GPT-6 name directly.

April 23, 2026: GPT-5.5 launches

Secondary coverage had associated the codename Spud with an expected GPT-6. OpenAI launched the model as GPT-5.5 and did not publicly confirm the GPT-6 label. Until OpenAI names a product, an internal codename or community label should not be treated as final.

May 20, 2026: an unnamed internal model produces original research

OpenAI reported that an internal reasoning model autonomously disproved the Erdős unit-distance conjecture, with external verification. The model was not named.

June 26 to July 9, 2026: GPT-5.6 rolls out in stages

GPT-5.6 Sol went through a 13-day limited preview restricted to roughly 20 trusted partners at the request of the US government, then reached general availability on July 9. This recent example suggests that OpenAI may test a frontier model with a small group before broad release.

July 21, 2026: OpenAI confirms a stronger pre-release model exists

In a security-incident disclosure, OpenAI stated that GPT-5.6 Sol and an "even more capable pre-release model" took part in internal cybersecurity evaluations. OpenAI has not established that this system was Astra or GPT-6, so this article keeps the identities separate.

July 26, 2026: a Washington preview is reported

Axios reported that Altman planned to demonstrate OpenAI's "most powerful AI yet" to policymakers that week, with an emphasis on complex work and agent teams. The report is evidence of a planned demonstration, not a GPT-6 product announcement or release date.

July 30, 2026: GPT-5.6 Terra and Luna prices fall

OpenAI reduced standard API prices for GPT-5.6 Terra to $2 input and $12 output per million tokens, and Luna to $0.20 input and $1.20 output. Sol pricing remained unchanged. These are current-model reference prices, not clues to Astra or GPT-6 pricing.

August 1, 2026: OpenAI names Astra

OpenAI published ten advances in mathematics and theoretical computer science produced by an internal version of Astra, calling Astra "our next major model." The post includes a paper, reasoning walkthroughs, and Lean certificates, but no product release date, ChatGPT rollout, model card, API model ID, or price. It also does not identify Astra as GPT-6.

September 3, 2026: GPT-6 Astra launches

OpenAI released GPT-6 Astra, published its model and safety documentation, and began a staged rollout. This event resolved the name and release-date uncertainty. Broad account access still follows the rollout schedule, and EvoLink availability remains a separate integration state.

The Rumors, Graded

Claim	Source type	Status
August 2026 launch target	X leakers (mid-July), relayed by tech blogs	Missed; official release was September 3
1.5M+ token context window	Same leak cluster	Missed; official context is 1.05M tokens
~10T-scale training run on Stargate (Abilene, TX)	Leaks; training location appeared in later reporting	Exact scale remains unverified
"GPT-6 released on April 14" with a "Symphony" MoE architecture	Recirculated fake-news pages	Debunked — nothing was released on that date under any name
Specific GPT-6 prices such as $2.5/$12	Third-party prediction articles	Missed; official standard list price is $10/$50
GPT-6 will be open-source	No source	No open-weight release was announced

Prediction markets were useful only as a temperature check, not evidence. The official September 3 release now supersedes those changing forecasts.

A Five-Minute Test for Any New GPT-6 Claim

When a post claims new GPT-6 access, a different model ID, a benchmark winner, or an EvoLink route, run this sequence before sharing it:

Open the original provider page: a blog quoting another blog is not the source. Look for an openai.com, developers.openai.com, or official Help Center record.
Match the exact name and wording: “next model,” “pre-release model,” and “GPT-6” are not interchangeable. Check whether the provider actually used the name.
Separate product, ChatGPT, and API claims: verify the model ID, endpoint, plan eligibility, regions, pricing, and rate limits independently.
Check the timestamp and page history: old predictions are often recirculated after their date has passed.
Look for reproducible access evidence: for API availability, documentation plus a successful authenticated request is stronger than a dashboard screenshot.
Keep track of what remains unknown: a confirmed model ID does not automatically mean general availability, stable capacity, or the same access on every provider.

This process intentionally favors being a few minutes slower over publishing a fictional price or route that later has to be retracted.

GPT-6 Astra Rollout Tracker by Channel

OpenAI released to a limited set of organizations on September 3 and broadened API and paid-plan access from September 4. Clouds, IDEs, and gateways listed the model on their own schedules. This table is updated when a channel changes state.

Channel	State (checked September 5, 2026)	Notes
OpenAI API	Live; rate limits published for Tiers 1–5, no Free tier	`gpt-6-astra`; Chat Completions does not support tool calling, tool calling needs Responses
ChatGPT	Rolling out to Plus, Pro, Business, and Enterprise; surfaced as "GPT-6 Pro" with per-plan weekly caps	Enterprise access is off by default at launch
Codex	Available; requires Codex CLI 0.153.0 or newer	Credit rates are 2.5x GPT-5.6 Sol
Microsoft Foundry (Azure)	Generally available since September 3, Global Standard and US Data Zone	No EU Data Zone and no Azure Batch at launch; Azure's pricing page still says prices are "in processing"
Amazon Bedrock	Not listed as of September 5	OpenAI's announcement names Bedrock; AWS model cards still end at GPT-5.6
GitHub Copilot	Generally available since September 4, gradual rollout	Pro+, Max, Business, and Enterprise plans
OpenRouter, Vercel AI Gateway	Listed September 4	OpenRouter also lists a separate `gpt-6-astra-pro` SKU
Cursor	Not listed as of September 5	Docs end at GPT-5.6
EvoLink	Callable	`gpt-6-astra`, 10% below OpenAI list, same key as GPT-5.6; see the API page

Two practical consequences:

Check the OpenAI account or API project that will actually run the workload; rollout eligibility can differ from public documentation.
Do not write "available on Bedrock" into a procurement document until AWS lists the model. Announcements name channels; listings prove them.

How to Call GPT-6 Astra on EvoLink

One field changes compared with a GPT-5.6 request. Tool calling requires the Responses API, so the example uses it:

curl https://api.evolink.ai/v1/responses \
  -H "Authorization: Bearer $EVOLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-astra",
    "input": "Summarize the release timeline of GPT-6 Astra in three bullets.",
    "reasoning": {"effort": "medium"}
  }'

temperature and top_p are not accepted, and none reasoning effort returns a 400. The GPT-6 Astra API guide covers the effort ladder, migration from GPT-5.6, and the 272K long-context price band.

GPT-5.6 Is Still the Cost Baseline

GPT-6 Astra lists at 2.5x the price of GPT-5.6 Sol. Keep GPT-5.6 as the cost baseline and promote Astra per workload:

Model	Input / Cached / Output (per 1M tokens)	Context / Max output
GPT-5.6 Sol	$4 / $0.40 / $20	1.05M / 128K
GPT-5.6 Terra	$2 / $0.20 / $12	1.05M / 128K
GPT-5.6 Luna	$0.20 / $0.02 / $1.20	1.05M / 128K

GPT-6 Astra uses the same 1.25x cache-write relationship, and requests above 272K input tokens are billed at 2x input and 1.5x output for the whole request. Compare the complete billing shape, not only the $10/$50 headline. OpenAI states that Astra uses fewer output tokens per task; whether that offsets the per-token price is workload-specific.

Both families run through one EvoLink key: see the GPT-5.6 page and the GPT-6 Astra page. The GPT-6 vs GPT-5.6 comparison turns this into a migrate-or-hold decision.

Developer Checklist for the First Astra Call

Use the exact ID gpt-6-astra on OpenAI and on EvoLink. There is no gpt-6 alias.
Send tool calls through the Responses API; Chat Completions does not support tool calling for Astra.
Remove temperature and top_p; map none or minimal effort to low.
Keep the model in configuration so switching between GPT-5.6 and Astra is one field change, not a migration.
Verify the returned model field and usage counters on the first request rather than trusting a dashboard label.
Keep a fallback route. Keep GPT-5.6 or Claude Opus 5 qualified; OpenAI's misalignment monitor can stop an Astra task mid-run.
Run a fixed set of real tasks on Astra and on GPT-5.6 before moving traffic; launch benchmarks are vendor evidence, not your workload.
Budget the whole billing shape: cache writes at 1.25x input, whole-request long-context rates above 272K, and Batch or Flex at 50% where latency allows.

What EvoLink Verified Before Activating the Route

EvoLink activated GPT-6 Astra after checking the request model ID and route mapping, the approved price and a real billing result, a successful authenticated request with the expected returned identity, and streaming, tool, structured-output, error, and long-context behavior. The API page is the source of truth for the current route state and price; this article records the release chronology.

Compare the Released Model Without Inventing a Winner

The Fable 5.1 vs GPT-6 Astra comparison uses published prices, cache economics, agent controls, and a matched evaluation contract without turning vendor benchmarks into a universal winner.

FAQ

When will GPT-6 be released?

GPT-6 Astra was released on September 3, 2026. OpenAI began with a limited set of organizations and said paid ChatGPT plans, API access, and AWS availability would expand over the following days.

Is there a GPT-6 waitlist I can join?

No public waitlist. OpenAI rolled access out by organization and plan from September 3 to 4. On EvoLink you can call gpt-6-astra with an existing API key.

How do I get access to the GPT-6 API?

Either through an eligible OpenAI account or API project, or through EvoLink, where GPT-6 Astra runs on the same OpenAI-compatible endpoint and key as GPT-5.6. The API page has the model ID, price, and code samples.

Is Astra GPT-6, and will the product be called GPT-6?

Yes. OpenAI officially released the product as GPT-6 Astra on September 3, 2026.

Will GPT-6 be free to use?

OpenAI's release statement names Plus, Pro, Business, and Enterprise plans. It does not establish general free-tier availability, and EvoLink access has separate commercial terms.

How much does the GPT-6 API cost?

OpenAI lists $10 input, $1 cached input, $12.50 cache write, and $50 output per million tokens, with 2x input and 1.5x output rates for requests above 272K input tokens. EvoLink prices GPT-6 Astra 10% below OpenAI list; current numbers are on the API page.

Is GPT-6 the model from OpenAI's security-incident report?

OpenAI's launch and safety materials now identify GPT-6 Astra as its first broadly deployed model at the Critical cybersecurity capability level. The older July disclosure did not name the internal pre-release system, so historical identity claims should still be attributed carefully.

Will GPT-6 have long-term memory?

Memory and personalization were pre-launch themes, but the published API model contract does not turn them into a standalone GPT-6 Astra memory API. Treat ChatGPT memory, API context, and data-retention policy as separate features.

How will GPT-6 compare to Claude Opus and Gemini?

Published specifications now support a direct facts table, but a winner still requires matched workload testing. See GPT-6 Astra vs Claude Opus 5 and Fable 5.1 vs GPT-6 Astra.

Should I wait for GPT-6 or build now?

Build now. GPT-6 Astra is callable, but it costs 2.5x GPT-5.6 Sol per token. Keep model selection configurable, run Astra as a challenger on a fixed task set, and promote it only where cost per accepted task improves.

Is GPT-6 Astra on Amazon Bedrock?

Not as of September 5, 2026. OpenAI's announcement names Bedrock as a channel, but AWS's OpenAI model cards still end at GPT-5.6. Azure Foundry lists it as generally available.

Sources

Evidence last reviewed September 5, 2026. Provider facts come from OpenAI, Microsoft, AWS, and GitHub documentation. GPT-6 Astra is callable on EvoLink; current price and route state live on the API page.

All Posts

#GPT-6 #OpenAI #Release Date #Release Watch #GPT-5.6