Yesterday Xiaomi open-weighted MiMo-V2.6 into the Artificial Analysis crown. Today the closed labs answered with a price war.

Ten days after Dario Amodei published We Must Pace the Frontier, Anthropic shipped Claude Opus 5.5 — Fable-class work, cheaper than the Opus it replaces, and the first model in a new 5.5 family. About ninety minutes later, OpenAI dropped GPT-6 Sol and GPT-6 Luna at half the GPT-5.6 promotional rates. xAI had already held Grok 4.7 at $2 / $6 on Monday. Three rate cards moved in 48 hours. Nobody waited for a regulator.

The slowdown, as a product calendar, lasted a week and a half.

Opus 5.5 is the first flagship that launched cheaper than the model it replaced

Anthropic’s pitch is blunt. Opus 5.5 “performs at the level of Claude Fable 5.1 on most work” and costs 40 percent less to run than Opus 5 on typical workloads. List prices fell 20 percent. The number that actually moves an agent bill is cache reads: $0.20 per million tokens, down 60 percent from Opus 5’s $0.50. Input / output landed at $4 / $20, against $5 / $25. Output is more than 30 percent faster. Fast mode exists at $8 / $40 if you want 2.5x speed and do not care about the multiplier.

The model ID is claude-opus-5-5. One million tokens of context, 128K max output. Live on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Sonnet 5.5 and Haiku 5.5 are promised in the coming weeks.

Vendor benches, with the usual grain of salt:

Anthropic itself says the gap versus Fable 5.1 is narrower in actual use than these scores imply. Treat the table as a cost-per-task argument, not a coronation. At default effort on FrontierCode, they claim Opus 5.5 beats Astra at roughly a fifth of the cost per task. On Terminal-Bench 4.0 it matches Astra for about 40 percent of the cost. That is the sentence an infra lead will screenshot.

Early-tester color, also from the lab: a 680,000-line migration in less than a day; a 200,000-line audit in under three hours versus more than 20 hours and 2.5x the tokens on Opus 5; an internal HAProxy C-to-Rust rewrite that finished in 9.5 hours versus 12 for Fable 5.1, at 51 percent lower cost. Those are anecdotes. They are also the kind of anecdotes that move seat counts.

Safety packaging is Fable-adjacent. Anthropic says Opus 5.5 is comparable to Mythos 5.1 on biology and cybersecurity, so it ships with Fable-class safeguards, external testing by METR and Frontier Design, and the strongest score yet on their automated behavioral audit. Reuters reported an 85 percent drop versus Opus 5 or Mythos 5.1 on a dedicated containment-bypass eval. Vetted orgs can apply to the Life Sciences Verification Program. Cyber verification expands in the coming weeks.

Four breaking API changes will eat a Saturday if you are still on Opus 5: thinking cannot be disabled, forced tool use returns an error, thinking blocks are tied to the model that produced them, and computer_20251124 is rejected on the Claude API and Google Cloud. The first three already applied to Fable 5.1. Read the migration guide before you flip a default.

Pro / Max / Team / seat-based Enterprise five-hour limits went up. Subscription users get a rate-limit reset they can bank.

Sol at $2 / $10. Luna at ten cents. Astra stays expensive on purpose.

OpenAI’s move is a different shape. Astra remains the flagship at $10 / $50. Sol and Luna are the workhorses, cut from the Astra generation, priced at half of GPT-5.6 promotional rates.

Sol now sits on the same list card as Claude Sonnet 5. Luna is the high-volume row: summarization, extraction, the jobs you run ten thousand times. OpenAI says caching and inference got cheaper and they are passing it through. Cached reads carry a 90 percent discount.

Availability is the messy part. Sol and Luna are in ChatGPT Work and Codex today for Plus, Pro, Business, Enterprise, and Edu. Both are in the API. Free and Go users can try Luna in the desktop app. They are not yet in Chat. Gradual ChatGPT rollout through the day, with the usual “to keep service stable” hedge.

On OpenAI’s internal factuality eval — de-identified conversations where users flagged mistakes — GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol, reaching Astra-level reliability at much lower cost, per TechCrunch. Alignment evals were updated the same day on the Astra system card for Sol and Luna; both are treated as High in cybersecurity and bio/chem, not High in AI self-improvement, with GPT-5.6-class safeguards.

There is no GPT-6 Terra in this drop. OpenAI has not said whether one is coming.

If you are actually routing this week: Opus 5.5 is the “pay less for Fable-shaped agent work” bet. Sol is the “keep OpenAI in the default and stop bleeding Astra prices on every loop” bet. Luna is the batch tax. Astra and Fable 5.1 stay on the hard jobs. That is a routing table, not a personality test.

A local box still wins when the loop is tight and the weights are yours. The closed labs just made the rented alternative less insulting.

The third cut was Monday, and it did not move

Grok 4.7 shipped September 21 at unchanged $2 / $6 — the rate Grok 4.6 already carried. That is 80 percent cheaper than Fable 5.1 and Astra on the sticker, and it undercut the room a day before Anthropic and OpenAI started marking down the middle of their own stacks. The 22nd was not a surprise. It was two labs refusing to let a third own the “good enough, cheap enough” row.

If you are buying a keyboard you will actually enjoy for twelve-hour agent sessions, the model card below the prompt matters less than the cost per merged PR. That number just moved on three providers.

CLOSEDQUORUM: four models vote. No human in the loop.

While the labs argued about dollars per million tokens, Cisco Talos published the first publicly documented Windows implant that treats commercial LLM APIs as command-and-control.

CLOSEDQUORUM is a 16.4MB 64-bit Go binary. After deploy, it queries up to four providers — DeepSeek, Qwen, Mistral, Google Gemini — and plurality-votes the next action. The system prompt, extracted from the binary: “You are an advanced malware strategist. Provide ONLY executable decisions.” The menu is constrained: steal, inject, persist, or move. Steal dumps LSASS, browser passwords (Chrome, Edge, Firefox), and crypto wallets (MetaMask, Exodus, Ethereum). There is no dedicated attacker C2 server. The endpoints are the same APIs thousands of legitimate apps hit every day.

Tie-break order, per Talos and later reporting: DeepSeek, then Qwen, then Mistral, then Gemini. If every model fails, the fallback is a consensus string with no handler — sleep and retry, not a default attack.

Caveats, because they matter. The public distribution build had placeholder API keys and a dummy Discord webhook. Talos did not observe a complete live loop and has no confirmation of in-the-wild victims. Artifacts tied the developer to carding-forum posts dating to 2025. This is a reference architecture, not a confirmed campaign. Treat “first autonomous AI C2” as “first publicly documented Windows implant that works this way,” which is already enough.

The hunting toolkit is the other half. CAIRN (Cognitive Artifact Intelligence Research Network) is open source on GitHub. Metadata-first: prompt templates, provider endpoints, API key prefixes, jailbreak strings. No download, no detonation. That is the useful artifact for anyone who actually runs a SOC, not another vibes-based “agents are scary” thread.

If your day job is keeping agents inside a box, a practical ML-systems book still beats a keynote. CLOSEDQUORUM is what happens when the decision space of one attack phase collapses onto a JSON schema and four HTTP clients.

Snorkel printed the data bill

Same day, quieter, and closer to why these models got cheaper to serve: Snorkel AI raised $350 million at a $3.5 billion valuation. Insight Partners and S32 led. Addition, Greylock, GV, and Wells Fargo stayed in. Valuation nearly tripled from the $1.3 billion Series D in May 2025.

The company shifted from labeling software to data-as-a-service last September. Annualized revenue is now about $375 million, up from roughly $20 million a year earlier — call it 18x. Coding data is one of the biggest demand lanes, alongside law and medicine. They expect to reach profitability this year. Frontier labs are not starving for tokens. They are starving for harder data and RL environments. Price cuts on the serving side and a 3.5 billion dollar data shop on the training side are the same week, not a coincidence.

What to actually do

• If Claude Code is your daily driver, try Opus 5.5 on one long migration before you rewrite the router. Cache-read cost is the lever. Watch the four breaking API changes.
• If you are on Codex / ChatGPT Work, Sol is the new default for “this would have been Astra last month.” Luna for batch. Leave Astra on the jobs that fail without it.
• Independent benches are not in yet. Vendor tables are a starting bid.
• If you hunt malware, run CAIRN against your existing corpus. You do not need a new incident to justify the metadata pass.
• Open-weight Monday still exists. Xiaomi’s MiMo-V2.6 did not get worse because two labs marked down APIs.

The pacing essay asked for time. The product orgs spent it on rate cards. That is not hypocrisy so much as the job: ship a safer, cheaper workhorse, put METR in the press release, and let public policy catch up “soon.” The malware authors did not wait either. They just stopped running their own C2.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.