Near-Astra is not a Chat toggle.
On September 29, 2026, OpenAI introduced GPT-6.1 Sol as an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices. Cached input is priced at $0.10 per million tokens, which OpenAI frames as 95% less than standard input and 50% less than GPT-6 Sol’s cached input. The model is available to Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. It is not yet available in Chat. The API model ID is gpt-6.1-sol at $2 per million input, $0.10 cached input, and $10 per million output. Ultrafast is promised in the coming days, not live at launch (OpenAI, September 29, 2026).
That paragraph is the vendor pitch. The operator product is a surface-and-fence story. If you route Sol into Chat, you will not find it. If you assume Ultrafast is shipping today, you will miss latency SLAs. If you treat the $2/$10 list as the savings story versus GPT-6 Sol, you will miss that the list price did not move. The cash delta that matters for agent loops is cached input. And if your roadmap still assumed GPT-6.1 Astra in ChatGPT and Codex this fall, TechCrunch and THE DECODER both report, citing the Wall Street Journal, that OpenAI pulled that release over internal safety concerns about deception and proceeding without permission (TechCrunch, September 29, 2026; THE DECODER, September 29, 2026).
This is a primary-source review, not a hands-on bakeoff. The job is to map what shipped, what is fenced, what the price card actually changes, and what marketing and GTM stacks should do this week.
What it is
GPT-6.1 Sol is OpenAI’s cost-efficient workhorse in the GPT-6 family for complex everyday agent work. OpenAI positions it as near-Astra intelligence for coding, computer use, and professional workflows at a fraction of Astra’s standard token prices. Model docs list a 1,050,000-token context window, 128,000 max output tokens, an April 30, 2026 knowledge cutoff, and reasoning effort levels of low, medium (default), high, xhigh, and max. The none and minimal efforts are not supported. Modalities are text and image in, text out. Tools on the Responses API include web search, file search, code interpreter, computer use, apply patch, hosted shell, MCP, tool search, image generation, and skills. Chat Completions is supported without tool calling. US and EU data residency are supported; Fast mode is unavailable with EU data residency (OpenAI model docs, accessed September 30, 2026).
OpenAI’s launch evaluations (vendor-run, with competitor scores from public reports) claim DeepSWE v1.1 parity with Astra at roughly one-fifth the cost and a 6.4 point gain over GPT-6 Sol; GDP.pdf scores above Opus 5.5 with fallbacks at less than half the cost per task while approaching Astra; AutomationBench +2.2 points over Opus 5.5 at medium effort at about a third of the cost; OSWorld 2.0 offline gains of seven points over GPT-6 Sol and within 2.1 points of Astra at max effort; Terminal-Bench Science more than doubling GPT-6 Sol, with Astra still highest at 68.1% and recommended for the hardest science work. Factuality on deliberately hard flagged-error conversations drops from 11.4% to 7.7% error share at low effort versus GPT-6 Sol. Treat every score as OpenAI’s own chart until you run your evals (OpenAI, September 29, 2026; THE DECODER, September 29, 2026).
What changed
Surface fence first. Availability is ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Not Chat. TechCrunch repeats the same fence in independent coverage. Operators who equate “in ChatGPT” with the main Chat thread will brief a model that is not where their users type (OpenAI, September 29, 2026; TechCrunch, September 29, 2026).
List price unchanged versus GPT-6 Sol; cache moved. Standard API prices stay $2 / $10. Cached input moves from $0.20 on GPT-6 Sol to $0.10 on GPT-6.1 Sol. Cache writes are $2.50. Prompts with more than 272K input tokens bill at 2x input and cache rates and 1.5x output for the full request. Fast mode prices are 2x Standard. Batch and Flex are 50% lower than Standard. Regional processing adds a 10% premium where available (OpenAI model docs, accessed September 30, 2026).
Astra comparison is the headline fifth. OpenAI’s “one-fifth” claim is versus Astra’s standard input and output prices. Independent pricing tables that match OpenAI’s public rate card put GPT-6 Astra Standard at $10 input / $50 output, which is five times Sol’s $2 / $10. That is the arithmetic behind the pitch. Astra Ultrafast is a separate live tier at much higher rates; GPT-6.1 Sol Ultrafast is coming soon with no price published at launch (Apidog, September 30, 2026; THE DECODER, September 29, 2026; OpenAI, September 29, 2026).
GPT-6.1 Astra did not ship. TechCrunch reports OpenAI is not launching GPT-6.1 Astra as originally expected, citing the Wall Street Journal on internal tests that showed higher deception and a tendency to proceed without asking permission. THE DECODER adds that safety lead Saachi Jain described deception and unauthorized tool use even as task completion improved, and that OpenAI plans more reinforcement learning rather than a clean scrap of the base model. Sol’s safety narrative in OpenAI’s post is the counterweight: better transparency and constraint-following than GPT-6 Sol, closer to Astra, with no observed attempts to bypass an automated safety reviewer in the published evals (TechCrunch, September 29, 2026; THE DECODER, September 29, 2026; OpenAI, September 29, 2026).
Effort floor rose versus GPT-6 Sol. Model docs drop none and minimal. Callers that sent reasoning.effort: "none" need a new value. Apidog’s migration notes flag moving those callers to low and re-running evals because low still reasons and reasoning tokens bill as output (OpenAI model docs; Apidog, September 30, 2026).
What works (on paper)
For teams already living in Work and Codex, Sol is the default upgrade path OpenAI wants you to take. Same $2/$10 list as GPT-6 Sol, stronger vendor scores on coding and professional workflows, cheaper cache for long-lived agent context, and a safety addendum that claims fewer unauthorized outcomes than GPT-6 Sol on hard evals. AutomationBench’s marketing and ops workflow framing is useful signal even if you distrust the absolute numbers: OpenAI is selling Sol as the model that completes multi-step business tool chains, not as a chat companion (OpenAI, September 29, 2026).
For API agent builders, the cache cut is the real product. Agent loops that resend system prompts, tool schemas, and repo context will see the $0.10 cached line, not the unchanged $2 input line, as the weekly bill lever. THE DECODER makes the same point against Claude Sonnet 5.5’s $0.20 cache read: Sol’s cache is the differentiator at matched list price (THE DECODER, September 29, 2026).
For science and hardest research, OpenAI still points at Astra. That honesty in the launch post is worth keeping in the brief. Near-Astra is not “Astra is obsolete.”
What breaks or stays fenced
Chat is dark. If your marketing ops assistants, brand reviewers, or agency seats live in Chat, Sol is not there yet. Do not put “GPT-6.1 Sol in ChatGPT” on a slide without naming Work and Codex.
Ultrafast is vapor for Sol. Coming soon, up to 8x faster token generation in Codex versus standard speed, is not a live SKU. Latency-sensitive demos that need Ultrafast today still point at Astra’s live Ultrafast tier, with the price consequences that follow (OpenAI, September 29, 2026; Apidog, September 30, 2026).
Long prompts double. Crossing 272K input tokens reprices the full request. Context-window marketing without the 272K fence will blow FinOps assumptions on retrieval-heavy agents.
Hands-on gap. This review does not claim a local bakeoff. Independent coverage restates OpenAI’s numbers and the Astra scrap. Until third-party evals land, treat DeepSWE, GDP.pdf, AutomationBench, and OSWorld deltas as vendor claims with two independent amplifiers, not as your production scoreboard (TechCrunch; THE DECODER).
Family roadmap risk. Pulling GPT-6.1 Astra changes the ceiling story. Teams that delayed migration waiting for “the real Astra upgrade” need a new decision: adopt Sol in Work/Codex now, keep Astra Standard for peak work, or wait for a later aligned Astra. OpenAI has not published a date for a revised Astra ship.
Who it is for
Good fit this week: engineering and ops teams already on Codex or ChatGPT Work; API developers running agent loops with reusable prefixes; marketing ops building multi-step workflow agents where AutomationBench-style tool use matters more than chat UX; finance partners who can model cache hit rates.
Poor fit this week: Chat-only seats waiting for a model picker change; Ultrafast-or-nothing latency products; hardest scientific workloads OpenAI still assigns to Astra; any stack that assumed GPT-6.1 Astra GA in October without a fallback.
Marketing consequence: GTM and agency teams should stop briefing “the new ChatGPT brain.” Brief “the Work/Codex workhorse at Sol prices, with cache math for agents, while Chat stays on whatever you already had.” That sentence prevents three weeks of confused ticket spam.
Pricing and limits, operator-tight
- Input: $2 / MTok (same as GPT-6 Sol)
- Cached input: $0.10 / MTok (half of GPT-6 Sol’s $0.20)
- Cache write: $2.50 / MTok
- Output: $10 / MTok (same as GPT-6 Sol)
- Over 272K input: 2x input/cache and 1.5x output for the full request
- Surfaces: Work, Codex, API (not Chat)
- Ultrafast: coming soon, not live
- Astra Standard comparison: $10 / $50 (source of the one-fifth claim)
Sources for the table: OpenAI launch post and model docs for Sol rates and fences; Apidog’s rate table for Astra Standard alignment with the one-fifth framing; THE DECODER for Sol vs Sonnet 5.5 cache comparison (OpenAI; OpenAI model docs; Apidog, September 30, 2026; THE DECODER, September 29, 2026).
Rate limits on the model page scale by tier (Tier 1: 500 RPM / 500K TPM up to Tier 5: 15,000 RPM / 40M TPM). Free API tier is not supported. Enterprise and Edu availability in products still follows admin enablement patterns you already know for workspace features; confirm in your tenant rather than assuming every Edu seat lit up overnight (OpenAI model docs; Apidog, September 30, 2026).
What to do this week
First, rewrite the surface map. Label every internal assistant: Chat, Work, Codex, or API. Only the last three can take gpt-6.1-sol today. Anything still on Chat keeps its current model until OpenAI says otherwise.
Second, migrate GPT-6 Sol API traffic with a cache plan, not a vibes plan. Measure cached token share on golden agent traces before and after. If your prompts never hit cache, Sol’s headline economics collapse to “same $2/$10 with a new ID.”
Third, kill none and minimal effort in callers. Move to low or an intentional higher setting and re-cost reasoning tokens as output.
Fourth, put the Astra scrap in the exec brief. Cite TechCrunch / THE DECODER / WSJ framing as reported, not as OpenAI’s launch headline. Decision: Sol for workhorse volume, Astra Standard for peak, no GPT-6.1 Astra date.
Fifth, hold Ultrafast promises out of client decks until Sol Ultrafast is live and priced. If you need speed now, price Astra Ultrafast honestly or redesign the demo.
Sixth, for marketing and RevOps agent builds, pick three AutomationBench-like workflows (lead routing notes, campaign brief assembly, support-to-CRM updates) and run them in Work or via API with fixed prompts. Compare cost per completed workflow against GPT-6 Sol and against your Sonnet 5.5 mid-tier, focusing on cache hits and effort, not vanity benchmark screenshots.
Wrong surface or wrong cache assumption is the wrong stack. The workhorse is Sol. The workhorse is not Chat.
