OpenAI launches GPT-6 Astra with 1.05M-token context, new pricing and reasoning tiers

OpenAI released GPT-6 Astra on September 3, 2026, according to a developer guide published on dev.to, rolling it out first to a limited set of enterprises before extending access to ChatGPT Plus, Pro, Business and Enterprise plans, the OpenAI API, Microsoft Azure and AWS Bedrock. OpenRouter also carries it under the name openai/gpt-6-astra. OpenAI describes the model as its strongest yet for software engineering, and dev.to notes it is the first model the company rates as critical for its cybersecurity capability, a classification that reportedly changes how it behaves through the API.

Model basics

The model identifier is gpt-6-astra. It accepts text and image input and returns text output, with a context window of 1,050,000 tokens, a maximum output of 128,000 tokens and a knowledge cutoff of April 30, 2026. It is reachable through chat completions, the Responses API and batch processing; real-time, the Assistants API and fine-tuning are not supported. Streaming, structured outputs, function calling, file search, web search, prompt caching and image input all work, alongside built-in tools such as computer use, web search, file search, a code interpreter and image generation. At rate-limit tier 5, dev.to lists 15,000 requests per minute and 40 million tokens per minute, and eligible API customers get zero data retention.

The Responses API is the primary interface for Astra and is required whenever tools are involved; chat completions remains usable for plain-text workloads.

Within ChatGPT, Astra usage draws on existing subscription allowances, and Pro, Business and Enterprise plans also receive a GPT-6 Astra Pro variant. Inside enterprise workspaces the model ships disabled, and an administrator has to enable it.

Reasoning levels and behavior

Astra supports five reasoning-effort settings: low, medium, high, xhigh and max. The none and minimal options from the GPT-5.6 generation are gone, and OpenAI's migration advice, as relayed by dev.to, is to map those settings to low and retest. That removal touches the cheapest tier of many pipelines, so dev.to suggests routing mechanical work to GPT-5.6 Terra or Luna and reserving Astra for harder requests.

Behavior shifts matter as well. The guide says the model asks for clarification more readily than its predecessors, is more sensitive to instructions embedded in skills and attached files, and defaults to lists and tables unless told otherwise. The max effort level is what OpenAI used for its launch benchmarks, and dev.to cites an Artificial Analysis run in which the first token took more than seven minutes to arrive at that setting.

Pricing

Standard pricing per million tokens is $10 for input, $1 for cached input, $12.50 for cache writes and $50 for output. Crossing 272,000 input tokens triggers long-context rates at roughly double: $20 input, $2 cached input and $75 output. Batch and Flex halve the standard prices, while priority mode doubles them and offers up to twice the processing speed.

GPT-5.6 Sol remains on a promotional $4 input / $20 output rate until at least November 21, 2026, against a listed price of $5 / $30, which places Astra at 2.5 times the promotional Sol cost for both input and output.

dev.to works through an example: an agent that reads a 200,000-token database snapshot once, reuses it as a cached prefix across 30 calls and emits 60,000 output tokens pays about $10.80 on Astra versus roughly $4.32 on promotional Sol, the same 2.5x ratio. OpenAI reportedly claims Astra is around 9% cheaper per task than Sol on Terminal-Bench 4.0 and uses about 65% fewer output tokens than Claude Opus 5 on Agents' Last Exam, so workloads that need fewer calls and tokens may close the gap; those that do not will simply pay 2.5x.

Migration from GPT-5.6 Sol

The migration involves three concrete changes. First, remove the sampling parameters temperature, top_p and top_logprobs, plus logprobs in chat completions and message.output_text.logprobs from the include list in Responses; OpenAI recommends deleting these fields from client code rather than relying on the API to ignore them. Second, convert any none or minimal effort settings to low. Third, replace prompt_cache_retention with prompt_cache_options.ttl.

Why it matters

Astra is a clean break in API design, not just a model-string swap: removed sampling knobs, no zero-reasoning tier and a restructured cache retention field all force code changes before migration. The 2.5x price premium over Sol's promotional rate pushes teams toward hybrid routing, keeping cheaper GPT-5.6 variants for routine work, while the doubled long-context rates above 272,000 tokens make context-window architecture a cost decision. The cybersecurity-critical classification is also a first for OpenAI and signals that capability-based restrictions may become part of future releases. All details here come from a single dev.to guide, so teams should verify against OpenAI's own documentation before committing migration code.