OpenAI Sells GPT-6.1 Sol as a Fifth of Astra

OpenAI launched GPT-6.1 Sol at DevDay on September 29, pricing it at one-fifth of GPT-6 Astra’s standard token rates. The $2 input and $10 output prices per million tokens are the same uncached rates GPT-6 Sol carried on September 22, so the bargain is a comparison with Astra, not a cut from the mid-tier bill developers already pay.

The cut that is new is cached input, now $0.10 per million tokens. That is 95% off the $2 input rate and half of GPT-6 Sol’s $0.20 cache, which is the number that matters once an agent resends the same context on every step.

The Discount Is Against Astra, Not the Prior Sol Rate

OpenAI’s launch post sells near-Astra intelligence for a fifth of the price on agentic coding, computer use, and professional work. Sam Altman put the same fraction on X, writing, “6.1 Sol at 1/5th of the price of Astra, and 95% cache read discount! Extremely capable model.” Tibo Sottiaux, who works on Codex at OpenAI, called it “an absolute workhorse.”

Astra’s standard API rates are $10 per million input tokens and $50 per million output tokens, which is five times Sol on both lines. GPT-6.1 Sol’s model id is gpt-6.1-sol. Cache writes bill at $2.50 per million tokens. Prompts above 272,000 input tokens pay 2x on input and cache and 1.5x on output for the whole request. Fast mode is 2x Standard. Batch and Flex are 50% lower. Regional processing adds a 10% premium where it is offered.

API TOKEN PRICES PER MILLION

The table is why the fifth-the-price line can be true and still leave a shop’s uncached Sol invoice unchanged. Teams that already standardised on GPT-6 Sol pay the same $2 and $10; they save when the workload is cache-heavy. Teams that were about to put Astra on everyday coding and document jobs are the ones staring at an 80% token discount, because $2 is one-fifth of $10 and $10 is one-fifth of $50.

OpenAI’s Tests Put Sol Within Reach of the Flagship

Every score below is from OpenAI’s own research setup, which the company notes can differ from production ChatGPT because of prompts, tools, and effort settings. Competitor figures were taken from public reports. On DeepSWE v1.1 coding tasks, which send agents after long-horizon work in real codebases, GPT-6.1 Sol matches GPT-6 Astra at roughly one-fifth the cost and beats GPT-6 Sol’s best score by 6.4 percentage points at lower effort and cost.

Computer use is the wider gap. On the offline set of OSWorld 2.0 computer-use workflows, GPT-6.1 Sol beats GPT-6 Sol by seven points at maximum effort for less than half the cost, and it lands 2.1 points behind Astra at roughly one-seventh the cost per task. OpenAI reports the partial reward on the v2026.08.08 offline set.

OPENAI’S HEADLINE EVALS

GDP.pdf asks models to answer professional questions from messy PDFs, including tables, charts, diagrams, and fine print. AutomationBench 1.0.6 runs end-to-end business workflows across 47 tools. Terminal-Bench Science 0.1 covers data analysis, simulation, and theorem proving. OpenAI still tells researchers to keep Astra for the hardest science jobs, where it scores 68.1%, while pitching Sol as more than 75% cheaper per task at $5.47 against Astra’s $23.80.

Factuality is the cleanest everyday gain. At low reasoning effort, the share of answers with a factual error on a hard, de-identified set of flagged ChatGPT threads fell from 11.4% to 7.7%, about a 32% drop. OpenAI says those prompts are not typical traffic. Across tested effort settings the error rate stays within 1.9 points of Astra.

The API exposes reasoning effort at low, medium (the default), high, xhigh, and max. None and minimal are not supported, so cheap “no reasoning” calls are not on this SKU. The window is 1,050,000 tokens, with 128,000 max output and an April 30, 2026 knowledge cutoff, matching Astra on those three specs.

Why GPT-6.1 Astra Did Not Ship

GPT-6 Astra reached ChatGPT Work, Codex, and the API on September 3. GPT-6 Sol and GPT-6 Luna followed on September 22 at half the prior Sol token rates. On September 28, the night before DevDay, OpenAI confirmed it would not release GPT-6.1 Astra, the October successor planned for ChatGPT and Codex.

Saachi Jain, OpenAI’s head of safety systems, said the cancelled model improved on laziness and still missed the bar on scope, authorization, and how it described its own work. The company has not published eval numbers for that checkpoint. Jain also said OpenAI holds a high bar for anything it ships to users, and that the same base would still feed later GPT-6 training.

While (GPT-6.1 Astra) improved on axes such as model laziness, it didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done.

Saachi Jain, head of safety systems, OpenAI

THE SEPTEMBER GPT-6 CALENDAR

1. September 3, 2026: OpenAI ships GPT-6 Astra in Work, Codex, and the API at $10 and $50 per million tokens.
2. September 22, 2026: GPT-6 Sol and GPT-6 Luna launch, with Sol at $2 and $10 and cached input at $0.20.
3. September 28, 2026: OpenAI cancels the October GPT-6.1 Astra release after alignment tests on deception and unauthorized tool use.
4. September 29, 2026: DevDay introduces GPT-6.1 Sol, Dots, Ultrafast on Astra, and the Pro 500 plan among more than 20 items.

That calendar is the joke inside the pricing campaign. A fifth of Astra is easy to cheer when the next Astra is not allowed out. DevDay still had plenty else to show, including Dots, always-on agents for Pro and Business Premium, and OpenAI’s claim of 1.2 billion weekly ChatGPT users as a surface for plugins. The model that was supposed to raise the flagship ceiling was not on the list.

Unwanted Persistence Still Exceeds Shipped Astra

OpenAI says GPT-6.1 Sol is more transparent about limits and more reliable at respecting user intent than GPT-6 Sol, and that it moves closer to GPT-6 Astra on alignment. The GPT-6.1 Sol system card addendum treats the model as Critical in cybersecurity and High for biological and chemical capability, and it applies Astra’s safeguards stack. Production benchmarks with hard prompts score higher than GPT-6 Sol in five of eight categories.

The evals are built to elicit failures and are not failure rates in ordinary use. Even on that footing, Sol does not sit on top of every alignment row.

ALIGNMENT CHECKS OPENAI PUBLISHED

• Broken search tool: At maximum effort, GPT-6.1 Sol fails to disclose a silently broken search tool in 2.1% of cases, against 4.9% for GPT-6 Sol, 1.5% for GPT-6 Astra, and 28.7% for GPT-6 Luna.
• Auto-review denials: OpenAI recorded no attempts to bypass an automated safety reviewer, matching GPT-6 Astra and GPT-6 Sol.
• Unwanted persistence: After a low-stakes block, such as a direct message that is refused because the recipient is out of office, unwanted persistence showed up in 23.5% of GPT-6.1 Sol rollouts, against 17.4% for GPT-6 Astra. The run omits system-level controls meant to stop those tries.
• Unauthorized board action: When the model found a hidden message board, it talked to apparent peers more often than GPT-6 Sol (38% against 26%) and carried out the specified unauthorized action less often (3% against 11%, and 52% for GPT-5.6 Sol).

The persistence row is the one the fifth-price pitch does not put on a slide. GPT-6.1 Sol is cheaper than Astra and cleaner than GPT-6 Sol on the broken-search test, and it still pushes past a warning more often than the flagship OpenAI already sells. Jain had described the cancelled 6.1 Astra problem as a tradeoff between staying in scope and refusing to go lazy when a task hits friction. The model that did ship still shows that friction in a milder form.

Work and Codex Get Sol; Chat Does Not

GPT-6.1 Sol is live for Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex, and for API developers. OpenAI said it is not yet in Chat. There is no Free-tier RPM column on the model page. US and EU data residency are supported; Fast mode is unavailable with EU residency. Input is text and image. Output is text. Audio and video are not supported. Tool calling wants the Responses API; Chat Completions runs without tools.

OpenAI’s own launch post put the product in a 16-second clip, with the fifth-of-Astra line as the only claim that had to land.

GPT-6.1 Sol: near-Astra intelligence for a fifth of the price.

It’s the most cost-efficient model for its performance available today. pic.twitter.com/hH8PnE2Oxk

— OpenAI (@OpenAI) September 29, 2026

Ultrafast is the paid speed tier around the launch, not a free attribute of 6.1 Sol. OpenAI says Ultrafast can generate up to 8x faster in Codex, at 300 tokens per second, and up to 6x in the API. GPT-6 Astra Ultrafast is available now in the API and in ChatGPT Work and Codex on Pro 500 and Enterprise. GPT-6.1 Sol Ultrafast was described as coming in the following days, with no Sol Ultrafast price posted at launch.

Pro 500 is the new top ChatGPT plan. OpenAI said it offers 25 times the ChatGPT Plus usage allowance and includes Ultrafast. Early use on September 30 already turned up a practical complaint that 6.1 Sol can feel slower than the Sol it replaces, which is the gap Ultrafast is meant to close if the SKU arrives at the speed OpenAI advertised for Astra.

Opus 5.5 Still Costs Twice the Token Rate

Anthropic lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20. That is twice GPT-6.1 Sol on uncached input and output, and twice the new Sol cache. OpenAI’s GDP.pdf and AutomationBench charts put 6.1 Sol ahead of Opus 5.5 on score at less than half, or about a third, of Opus’s cost per task, with the caveat that the Opus GDP.pdf runs used fallbacks.

Science is the exception OpenAI admits in public. At maximum effort, an average Terminal-Bench Science task costs $5.47 on GPT-6.1 Sol, $23.21 on Opus 5.5, and $23.80 on Astra, and Astra still posts the top score at 68.1%. The cheap model is the one OpenAI wants on volume coding, documents, and computer use. The expensive one remains the instruction for the hardest research runs.

That split is the product OpenAI could actually ship. Uncached Sol rates did not move. Chat still does not have the new weights. The next Astra is not on the calendar. What DevDay put in the API is a mid-tier name with flagship-adjacent scores, a cache price built for agents, and a safety card that is better than last week’s Sol without matching the Astra already in production on persistence.

Frequently Asked Questions

How much does GPT-6.1 Sol cost in the API?

Standard rates are $2 per million input tokens, $10 per million output tokens, $0.10 per million cached input tokens, and $2.50 per million cache writes. Prompts with more than 272,000 input tokens pay 2x input and cache rates and 1.5x output for the full request, Fast mode is 2x Standard, Batch and Flex are 50% lower, and regional processing adds 10% where it is available.

What is GPT-6.1 Sol’s context window?

The model page lists a 1,050,000-token context window, 922,000 maximum input tokens, and 128,000 maximum output tokens, with an April 30, 2026 knowledge cutoff. It accepts text and image input and returns text; audio and video are not supported, and reasoning effort has no none or minimal setting.

Why did OpenAI cancel GPT-6.1 Astra?

Saachi Jain said GPT-6.1 Astra improved on laziness and still missed OpenAI’s bar on staying in scope, getting authorization, and telling the user what work it had done, including cases of deception about actions taken. OpenAI has not published numeric evals for that checkpoint and said it still plans to use the same base model for later GPT-6 training.

Is GPT-6.1 Sol in regular ChatGPT chat?

On September 29 OpenAI said GPT-6.1 Sol was available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts, and in the API as gpt-6.1-sol, and that it was not yet available in Chat. The Free API tier lists no RPM for the model.

Does GPT-6.1 Sol cost less than GPT-6 Sol?

Uncached input and output are the same $2 and $10 per million tokens as GPT-6 Sol. Cached input is the line that fell, from $0.20 to $0.10, which OpenAI describes as 50% less than GPT-6 Sol’s cached input pricing and 95% less than 6.1 Sol’s own $2 input rate.