The AI Catchup - September 29, 2026
Welcome back. OpenAI held DevDay 2026 today and shipped more than 20 announcements. Three of them change what you can do this week: a cheaper model that codes like the flagship, an agent that keeps working when you close the tab, and a new top plan that sells speed. One more changes what you pay: by the figures OpenAI's Tibo Sottiaux gave, the $200 Pro plan now buys half the Codex usage it did.
We read every launch page, help article, and model doc, and fact-checked each claim in our full write-ups. Here is what changed, what it costs, and what to do about it.
Let us get into it.
At a Glance
| What changed | What you can do now that you could not before | Can you use it today? |
|---|---|---|
| GPT-6.1 Sol replaces GPT-6 Sol at the same price | Get Astra-level coding scores for $2 and $10 per million tokens instead of Astra's $10 and $50 | Yes, as gpt-6.1-sol in the API, and in ChatGPT Work and Codex on paid plans |
| dots, always-on agents in ChatGPT | Hand an agent a standing goal and let it keep working between conversations on its own cloud computer | Rolling out to Pro (outside the EEA, Switzerland, and the UK) and Business Premium |
| Pro 500 and Astra Ultrafast | Run GPT-6 Astra up to 8x faster in Codex, about 300 tokens per second, with 25x Plus usage | Yes, subscribe on the web; Ultrafast is also in the API |
| Pro 200 allowance changes | Nothing new: Pro 200 reopens to new subscribers at a smaller allowance, reported as 10x Plus instead of 20x | Existing subscribers keep the old allowance through October 29 |
GPT-6.1 Sol: Astra-Level Coding at Sol's Price
Last week, the best DeepSWE score OpenAI sold was GPT-6 Astra's 74.1%, at $10 and $50 per million tokens. Today GPT-6.1 Sol scores 75.2% at high effort for $2 and $10, about $0.65 a task against Astra's $4.43. OpenAI's own text calls that a match; its chart puts 6.1 Sol slightly ahead. Either way, the price of flagship-grade coding just fell by four-fifths.
It is a one-week-old model's replacement, and it beats GPT-6 Sol's best score on all six benchmark charts OpenAI published. Cached input halves to $0.10 per million tokens, a tenth of Astra's $1, which matters most for agents that resend a long prefix every turn.
| Benchmark (best score) | GPT-6.1 Sol | GPT-6 Sol | GPT-6 Astra |
|---|---|---|---|
| DeepSWE | 75.2% ($0.65) | 68.8% ($2.74) | 74.1% ($4.43) |
| OSWorld offline | 71.4% ($1.27) | 64.4% ($3.37) | 73.5% ($9.44) |
| AutomationBench | 36.1% ($0.30) | 33.2% ($0.27) | 41.4% ($1.73) |
| GDP.pdf | 32.0% ($0.35) | 28.0% ($0.35) | 32.2% ($1.91) |
| Terminal-Bench Science | 57.0% ($5.47) | 27.6% ($12.18) | 68.1% ($23.80) |
Costs are OpenAI's per-task estimates, and every score is vendor-reported. Astra still earns its price in three places: computer use (2.1 points ahead on OSWorld offline), the hardest business workflows (5.3 points on AutomationBench), and long scientific tasks (11.1 points on Terminal-Bench Science).
Pick your effort level from the charts, not by instinct. On DeepSWE, 6.1 Sol peaks at high; xhigh and max both drop to 71.9% and cost up to 2.4 times as much. The default, medium, already scores 73.0% for $0.42 a task, above GPT-6 Sol's best at any setting. Go higher only for business workflows, factual answers, computer use, and science.
Two API changes will break existing code. The none effort level is gone, and Chat Completions no longer supports tool calling, so agent loops belong on the Responses API. Against Claude, Sonnet 5.5 costs the same per token, and the two labs publish no benchmark in common, so run both on your own tasks. A GPT-6.1 Sol Ultrafast is coming; OpenAI has not given a date or price.
dots: An Agent That Keeps Working Between Conversations
Until today, ChatGPT and Codex agents ran when you asked them to. A dot takes a standing goal and keeps making progress between conversations, on its own cloud computer, then brings results back for your review. It runs on GPT-6 Astra, remembers context from the apps you connect, and can hand coding work to Codex by creating cloud tasks in your existing Codex environments.
You decide how much it does alone. For some actions, you set custom rules with one of four behaviors: take action without asking, take action if pre-approved, ask first, or hand off to you. OpenAI's own caveat: your dot can make mistakes, including when following your rules.
Who gets it. Dots are rolling out gradually to Pro and Business Premium users aged 18 and older. Pro access excludes the EEA, Switzerland, and the UK at launch; Enterprise is a beta that is off by default. You create one on the ChatGPT desktop app or desktop web, not on mobile.
What it costs. Nothing extra for now. For the next month, dots usage won't count toward eligible plan allowances, and OpenAI has not said what happens after. Use the month to measure how much an always-on Astra agent actually consumes before you size a plan around it.
One thing to know before you connect anything sensitive: disconnecting an app does not delete what your dot already learned from it. Deleting that means resetting your dot, which wipes its conversations, memories, and scheduled tasks.
ChatGPT Space launched alongside it as a shared home for Pages, files, and project work that teammates, ChatGPT, and your dot can build on, with a Meetings plugin (beta, macOS desktop) that saves summaries and action items there and deletes the audio. Our full dots explainer compares dots with Codex, workspace agents, Meta's Muse, and Claude Code Routines.
Pro 500 and Ultrafast: OpenAI Now Sells Speed
Codex already had Fast mode. Ultrafast is a new, faster service tier for GPT-6 Astra: up to 8x Standard token speed in Codex, about 300 tokens per second, and up to 6x in the API. Among Pro plans, it comes only with the new $500-a-month Pro 500, which also carries OpenAI's highest allowance, 25 times ChatGPT Plus. Buying credits on Pro 100 or Pro 200 does not unlock it.
| Plan | Monthly price | Codex and ChatGPT Work usage | Astra Ultrafast |
|---|---|---|---|
| Plus | $20 | Baseline | No |
| Pro 100 | $100 | 5x Plus | No |
| Pro 200 | $200 | 10x Plus for new subscribers (was 20x) | No |
| Pro 500 | $500 | 25x Plus | Yes |
Speed is not free inside the plan. Ultrafast draws on included usage first, at 8x the Standard rate, then on credits at 6x. Run Astra on Ultrafast all week and Pro 500's 25x stretches about as far as 3x Plus of Standard Astra. OpenAI also notes the 8x is token generation speed, not task time: a run dominated by tests and tool calls will not finish 8x sooner. In the API, Ultrafast Astra costs $60 input and $300 output per million tokens, six times Standard, set with service_tier: "ultrafast".
The Pro 200 change is the real news for most readers. Every Pro tier now works out to $20 per Plus-sized allowance, the same rate as Plus itself. Under the old Pro 200 terms, 20x for $200 was $10. OpenAI's pages confirm Pro 200's allowance is lower and that eligible existing subscribers keep their previous allowance through October 29, 2026; the 10x figure comes from OpenAI's Tibo Sottiaux and The New Stack.
Our call: pay $500 if you already use more than half of the old Pro 200 allowance, or if waiting on Astra in interactive work (pair programming, debugging, UI iteration) is your bottleneck. Stay on Pro 200 if 10x Plus covers you, or you mostly run GPT-6.1 Sol, which has no Ultrafast yet. Check your September usage dashboard before October 29, not the launch-day headline. For comparison, Anthropic still sells Claude Max 20x at $200, and Claude Code's fast mode is paid from usage credits only. The full plan breakdown has the Claude comparison.
Also From DevDay
These come from OpenAI's DevDay recap:
- The Codex CLI takes voice. You can start and steer tasks by voice, and a new
/agentsview tracks multiple tasks at once. Available on all plans. - Code review arrives in the ChatGPT desktop app, with summaries, diffs, and automatic first-pass reviews in the cloud before you comment on a GitHub pull request or GitLab merge request. Available on all plans.
- The Agents API supports computer use and brings Codex's multi-agent capabilities, tool search, and context compaction to your own application.
- Sign in with ChatGPT spends your plan in other tools. Plus and Pro users can use their allowance across 16 partners, including Cognition's Devin, Notion, and Vercel, and cap how much each can use.
Quick Hits
- Claude Sonnet 5.5 is out at Sonnet 5's $2 and $10 per million tokens, with output more than 30% faster and up to 30% lower cost per task by Anthropic's figures. It is already in Cursor. What changed.
- Claude Code can build evals and hillclimb against them.
/claude-api build-evalcreates an evaluation in your codebase, and/claude-api hillclimbtries one change at a time against held-out tests. How it works. - Cursor Rollouts watches your deploys. A Teams and Enterprise bot follows a change from pull request to production and reports regressions from logs, metrics, and traces. The details.
- Claude's plugin directory takes submissions. Paid-plan developers can submit remote MCP connectors or plugin bundles for review. How to submit.
Ship It This Week
Move coding agents from gpt-6-sol to gpt-6.1-sol at high effort. Replace any none effort setting and move tool calls from Chat Completions to the Responses API first.
Rerun your Astra coding evals on GPT-6.1 Sol. On OpenAI's DeepSWE chart it matches Astra for about a seventh of the cost per task. Keep Astra for computer use and science.
If you have dots access, give it one recurring job this month. Usage is off the meter, so it is the cheapest time to learn what it costs.
Decide on Pro 200 vs Pro 500 before October 29. Pull your September Codex usage. If it tops half of the old allowance, the $500 tier is the only self-serve plan that leaves you more room than you had.
Account and Billing Notes
- Pro 200 grandfathering. If your Pro 200 subscription was active at the eligibility cutoff or in the seven days before, you keep the previous allowance through October 29, 2026, then move to the lower one. Keeping it does not add Ultrafast. Details.
- Claude bills some blocked requests again. Anthropic resumed billing for pre-output refusals in three documented safeguard categories. Which ones.
That is it for this issue. The theme of DevDay was that the flagship is no longer the default: GPT-6.1 Sol does most of Astra's coding work for a fifth of the token price, and OpenAI now charges extra for Astra's speed rather than its brains. If a teammate is still paying Astra prices for everyday coding, forward this.
Until next week, stay caught up.