OpenAI’s DevDay answered the cancellation of its flagship this morning with the opposite of a flagship: GPT-6.1 Sol, a model its own benchmarks put within reach of Astra at a fifth of the price, on the same $2/$10 card Anthropic’s Sonnet 5.5 now carries; and Dots, an always-on consumer agent running on the Astra that passed the safety review, not the one that didn’t. Stack the pair against the flagship cancelled yesterday, the prospectus Reuters put on the record last night, and the up-to-$84.5 billion SpaceX compute line reported out of it this morning, and the week’s shape reads structural rather than accidental. The two labs with the most at stake in the frontier race are rationing the flagship tier and flooding the $2 tier. Safety reviews and balance sheets are steering capability toward the mid-range, and the agents that run on it are being sold as subscriptions.
The near-frontier shipped at the mid-tier price, and the cache read is the tell
What happened. At DevDay’s keynote OpenAI released GPT-6.1 Sol: on its own numbers it ties Astra on the DeepSWE v1.1 agentic-coding benchmark at about a fifth of the cost, lands 2.1 points behind it on the OSWorld 2.0 computer-use suite at roughly a seventh of the cost, and more than doubles its predecessor’s Terminal-Bench Science score. API pricing is $2 per million input tokens and $10 per million output — the same card GPT-6 Sol has carried since the September 22 pair — with cached input at $0.10, half of Sonnet 5.5’s $0.20. It is live today in ChatGPT Work, Codex and the API as gpt-6.1-sol for Pro, Business, Enterprise and Edu, though not yet in regular chat, with an Ultrafast variant due in the coming days. The top-shelf Astra the calendar expected stays off it.
Why it matters. Yesterday’s breaking note asked whether DevDay would produce a replacement for GPT-6.1 or go quiet; the answer is a replacement priced as a mid-tier, which is a decision about where a frontier lab routes its best current capability, not about how good the model is. The operative detail for anyone paying per token is the cache line — $0.10 against Sonnet’s $0.20, halving the charge that dominates the bill of long-running agents, the same cache-read price war that moved there last week. OpenAI just matched Anthropic’s workhorse price and undercut the line that actually accumulates when an agent runs for a day. Read next to the Sonnet 5.5 shipping Monday, the two leaders have converged on an identical $2/$10 workhorse with the edge decided exactly where sustained agent workloads bill. Capability is being routed around safety by pricing: the model that cleared review gets flooded at the cheap tier, and the one that didn’t stays off the calendar.
Source: platform.openai.com, the-decoder.com
The always-on agent shipped on the model that passed, priced as a subscription
What happened. OpenAI also launched Dots, always-on agents that run around the clock on their own cloud computers — browser, terminal, file storage — reachable through ChatGPT, Slack and Microsoft Teams, powered by the shipped GPT-6 Astra base. Plugins connect a Dot to more than 4,000 apps; each account gets one at launch, with teams of Dots planned. In background “proactive research” mode the agent only uses read-only tools, actions that touch accounts pass an automated rule check, and password changes stay human-only. Dots roll out today to ChatGPT Pro and Business Premium, with an admin-enabled beta for Enterprise, Edu and Healthcare and a restricted launch in Europe. The same keynote reopened OpenAI’s $200 Pro plan with a thinner allowance and added a $500-a-month Pro Max tier.
Why it matters. Dots is OpenAI’s answer to Meta’s Muse — the always-on consumer agent reported to have drawn half a million users in its first week — and it arrives with the containment week’s vocabulary already installed: read-only fallback, permission gates on account-touching actions, humans-only on the sensitive stuff. That is a softer guarantee than the off-host watchdog NVIDIA shipped Monday; here the rails live in the product’s own rules engine and its training-data choices, which is exactly where a practitioner should look before trusting a background agent with their accounts. And the pricing shape tells you where the agent layer is heading: OpenAI is selling the always-on category as a $200/$500 subscription rather than per token, and it deliberately put the consumer surface on the model that cleared safety review while the questionable one stays off the calendar. The agent is the product; the model underneath it has become infrastructure.
Source: the-decoder.com
Anthropic’s half-trillion-dollar buildout is the floor under the price war
What happened. The prospectus Reuters reviewed last night already put about $4.6 billion of revenue (up ~12x), a roughly $42 billion net loss, and about $518 billion of committed cloud-and-compute obligations on the record, most of it described as deals that largely cannot be cancelled. Today The Information reports the filing also discloses up to $84.5 billion in SpaceX compute agreements, nearly double prior estimates.
Why it matters. The commitment stack is the reason the price war has a floor you can build against. Most of that $518 billion gets spent whether or not anyone buys tokens, and once the capacity is sunk the marginal economics invert: the winning move becomes volume at the low end — exactly what Sonnet 5.5’s cheaper-per-task card and Sol’s matching $2/$10 line are doing. Anyone budgeting for the mid-tier to get more expensive again is betting against the balance sheet of the lab that is about to go public on these very commitments. Read with the flagship cancellation and the week is consistent in writing: the leaders are pricing risk differently — Anthropic in an S-1 that lists existential disclaimers as line items, OpenAI as a safety review that cancelled a release.
Source: reuters.com, theinformation.com
The Rest
- The White House AI summit opened with America.gov and a rename order — Trump signed the executive order launching the AI-powered America.gov portal and said he would sign another renaming “artificial intelligence” as “Superintelligence” in government documents; still no named czar. The vocabulary move is the week’s “define the term, control the test” thread scaled to a government. politico.com
- The joint US-China AI venture is now explicitly ruled out — Trump said jointly developing AI with China would hurt US companies, closing the bilateral thread Sunday’s zeitgeist flagged, the same week the race between the two arrived without agreed rules. bbc.com
- OpenAI answered the decision-layer race with a Decisions API on Luna — a direct taker on the thread tracked since TypeSafe’s Jev on the 19th: choosing as an API endpoint, now from OpenAI too. thenewstack.io
- Anthropic’s S-1 flags uncertain legal risk from rogue agents — the lab that just shipped Sonnet 5.5 warns in writing that autonomous-agent harm may expose it to uncharted liability, the securities-filing echo of the government-breach accountability arc. reuters.com
- Mercury News: stop blaming rogue bots, hold the human developers accountable — a Bay Area voice putting the week’s accountability thread where it belongs, the paragraph the containment arc has been missing. mercurynews.com
What I’m watching
Whether the Ultrafast Sol and the rest of the keynote land anything at the top shelf, whether a named czar or hard executive action comes out of the summit day, whether Anthropic’s EDGAR accession turns the leaked prospectus into an official S-1, and the first real deployments of Dots in background mode — where its read-only rules, not its demo, get tested at load. Also carrying the reopened $200 Pro plan’s thinner allowance as the honest test of whether always-on agents are being priced to run continuously.