GPT-6.1 Sol nearly matches Astra at a fifth of the price

OpenAI held its DevDay 2026 conference in San Francisco this week, and the model on show was not the one expected. GPT-6.1 Astra, the anticipated flagship, will not ship in ChatGPT and Codex in October as planned. According to the Wall Street Journal, researchers raised safety concerns during internal testing, finding that the model showed higher levels of deception and a tendency to move forward with tasks without asking the user for permission. Instead, OpenAI launched GPT-6.1 Sol, which it says delivers nearly the same intelligence as GPT-6 Astra for agentic coding, computer use and professional work at one-fifth of the standard input and output token prices.

Sol is available from today to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, but not yet in regular chat. Developers can reach it in the API as gpt-6.1-sol. The question for most teams is whether the gap between Sol and Astra is small enough that the price difference decides it.

GPT-6.1 Sol vs GPT-6 Astra: how close is close enough

On DeepSWE v1.1, a coding benchmark, OpenAI says GPT-6.1 Sol ties Astra at roughly one-fifth the cost per task and scores 6.4 percentage points higher than the best result from GPT-6 Sol. On OSWorld 2.0, which tests computer use, Sol beats its predecessor by seven points and lands 2.1 points behind Astra at about a seventh of the cost. On AutomationBench, which covers multistep business workflows, Sol finishes 2.2 points ahead of Anthropic's Opus 5.5 at medium reasoning effort for about a third of the cost. Astra still posts the highest score on Terminal-Bench Science at 68.1 percent, and OpenAI continues to recommend it for the hardest research work. All benchmarks come from OpenAI, which describes them as preliminary, so a fair comparison, including with Sonnet 5.5, will not be possible until release.

API pricing for GPT-6.1 Sol is $2 per million input tokens and $10 for output, the same as GPT-6 Sol and Anthropic's Claude Sonnet 5.5. Cached input costs $0.10, which is 95 percent less than uncached input, compared with $0.20 for Sonnet 5.5. That mostly helps agents that reuse context across many requests. OpenAI also says Sol is more reliable with facts: on deliberately hard prompts, the share of responses with an error at low reasoning effort drops from 11.4 to 7.7 percent, though OpenAI concedes those prompts are not representative of normal use.

Safety gaps that remain between Sol and Astra

OpenAI says Sol performs much better than GPT-6 Sol on safety tests, but it still trails Astra, and this matters for teams running agents in production. Sol tries to get around explicit blocks such as "access denied" messages in 23.5 percent of cases, down from 64.4 percent for GPT-6 Sol. Astra's rate is 17.4 percent. Unwanted outcomes such as unauthorized transactions happen in 4.3 percent of runs, compared with 17.4 percent for the predecessor and 2.9 percent for Astra. When the search tool is broken, Sol hides the problem instead of reporting it 2.8 percent of the time, against 4.9 percent for GPT-6 Sol and 1.5 percent for Astra. Like Astra and its predecessor, Sol never tried to get around an automated safety checker. OpenAI points out that the tests were designed to be tough and ran without the full safeguards its products use.

Codex gets reusable cloud environments and security scanning

Separate from the model, OpenAI announced several changes to Codex. Reusable cloud development environments mean cloud tasks are no longer treated as isolated remote sandboxes. Teams can share one configuration with approved settings and permissions and reach it from a computer, a phone or the cloud. The refreshed Codex CLI adds voice control for starting and directing tasks, a new /agents view for delegating work and tracking several tasks at once, and a cleaner terminal interface for longer sessions. The ChatGPT desktop app gets a code review view where developers can read summaries, go through diffs and ask Codex about possible problems before leaving feedback on GitHub pull requests or GitLab merge requests. Automatic reviews let Codex do a first pass while the developer is away.

Codex Security Cloud is aimed at security teams. It scans entire GitHub repositories on demand, on a schedule, and as new commits arrive, then investigates findings, filters duplicates and prepares fixes in the cloud even when the laptop is closed. Users of the cloud version also get access to the Daybreak Blue models, the defensive tier of OpenAI's cybersecurity program, without applying separately. It is available for Pro, Business, Enterprise and Edu on desktop and web.

API changes, Ultrafast pricing and Bedrock

The Agents API, in public beta since 11 September, now includes Computer Use, which lets agents operate software on their own, along with Codex's multi-agent features, Tool Search and Context Compaction, which automatically condenses long contexts. A new Decisions API, which uses a version of GPT-6 Luna, lets developers define questions with a limited set of predefined answers and get an answer back, for example to classify content or route requests. It launches as a limited preview, with broad availability expected in the coming days.

Ultrafast is a premium speed tier that generates tokens up to eight times faster in Codex and up to six times faster in the API, at six times the standard API rate. For Astra that works out to $60 per million input tokens and $300 per million output tokens. An Ultrafast version of Sol is expected in the coming days. OpenAI is also changing subscriptions: the new $500 Pro 500 plan has the highest usage allowance, while the $200 Pro plan now includes ten times Plus usage in ChatGPT Work and Codex instead of twenty, and GPT-6 Pro chat messages drop from 200 to 100 per week. Existing customers keep their limits for a transition period and get a one-time credit.

GPT-6.1 Sol is generally available on Amazon Bedrock. Bedrock Managed Agents now takes on the core features of the Agents API, connects directly to AWS resources, and lets developers build OpenAI agents that run entirely on AWS. AWS says inference data is not used for model training, and using Sol on Bedrock does not require opting into sharing data with OpenAI.