The Product Desk

The model in the headline is not the model on the price sheet

Meta's launch contains two products. Muse Wonder Full produces the headline numbers: 57.5 on SWE-Marathon against 46.3 for Claude Opus 5.5. That score is vendor-reported, on Meta's own benchmark. Artificial Analysis, which scores models independently, shows a tighter race: Full at 56 against Opus 5.5 at 58. Full has no API price, and only Muse Code subscribers can use it. The model you can call through the API, standard Muse Wonder, scores 52. It costs $4 per million input tokens and $20 per million output tokens.

The figure that matters for your margins is cost per task, not cost per token. Artificial Analysis puts standard Wonder at $4.52 per benchmark task. That is nearly triple the $1.60 of Meta's own Spark 1.3, which scored 48 six weeks ago. Four index points cost roughly 2.8 times as much per task. That trade pays off for some features and wrecks the economics of others. Only your own task mix tells you which features are which.

The cheap end keeps getting cheaper

While the top tier gets more expensive, the bottom is falling in two ways. Turing Post reports that Anthropic's Claude Haiku 5.5 runs about 75% cheaper on average than Haiku 4.5. It targets classification, summaries, database queries and coding subagents. Anthropic also halved Sonnet 5.5's cache-read price.

A new category, decision models, also arrived within a few weeks. These models pick from a fixed set of options and return a probability for each, instead of writing text. Microsoft-Decision-1 charges $0.042 per million input tokens, with output free, and reports 85ms median latency. SubQ Snap charges $0.05 per million input tokens. AI Breakfast notes that Microsoft ran its own benchmarks and that H2O.ai disputes Microsoft's latency comparisons against competitors.

ByteByteGo covers the most extreme case of using the wrong architecture. In a public Sudoku test, a panel of frontier LLMs solved 2% of puzzles and spent about $11,000 on API calls. Kona, an energy-based model roughly a thousand times smaller, solved 96.2% for $4 of GPU time. These figures come from the vendor, and they compare GPU time against retail API prices, so we think the ratio overstates the real gap. The direction still holds. When a machine can check whether an answer is right, a general-purpose LLM may be the slowest and most expensive way to get it.

Where the evidence disagrees

The 14% figure from The Information does not say whether 'usage' means tokens, requests or spend. OpenRouter's users also lean toward developers who switch models easily. Meta argues the other way. Its Active Focus feature keeps the model working through obstacles on long tasks, which is a case for paying frontier prices on long-running agent work. Martin Peers of The Information suggests Meta's biggest gain may be lower serving costs for itself, since consumers don't need the most advanced model for everything. We think both readings are true at once: frontier models for a narrow set of long tasks, cheaper tiers for volume. The Information AM also carries a headline saying Microsoft cut its internal Claude spending by a third. There is no detail behind it, so treat it as unconfirmed.

When four labs claim the frontier in six weeks, the claim stops carrying information. Your cost per completed task still does.

The smart move

Stop planning your roadmap around launch dates. Map each AI feature to the cheapest tier that meets its quality bar. Test new flagships on a fixed evaluation schedule, not whenever a launch post appears. Then decide on purpose whether to keep the margin you recover, reinvest it, or pass it to customers.

What the demo proved, and what it didn't

Latent.Space's account of the AI Engineer New York workshop describes one demo, a stock watchlist scanner. In the scan phase, each ticker goes to its own agent, and no agent sees the others' work. In the review phase, agents check those findings before a report is written. CJ Avilla of Anthropic said the flow was 'much more effective at finding all of the details' than a single agent. He also said it cost 'much more.' Anthropic shared no cost figures, latency numbers or benchmarks. The demo shows a pattern, not proof, and the return-on-investment math is left to you.

Three limits shape what you can build today:

• The model can't see the budget. You set a session budget, but Claude doesn't know it while it works. Harrison Stall joked that telling models their budget might make them 'a little bit anxious.'
• Agent count rests on defaults. The number of agents is held down by 'evals and lower defaults.' Stall conceded the point when asked whether a ticker scan could spawn hundreds of subagents.
• Agents can't talk directly. Subagents share a filesystem, but the API has no way to tell one subagent to message another. Workflows where agents need to negotiate with each other are out for now.

You can still give Claude your own specialist subagents through the agent configuration. A predefined team is now optional rather than gone. Claude Console also draws a graph showing what the lead agent passed to each subagent at every handoff. That helps your engineers debug. It is not a review experience for your users.

Spending caps are becoming standard in agent platforms

Flows aren't the only launch that leads with cost control. Turing Post reports that Google Cloud's Gemini agent includes spending controls and model routing. Cohere North 2 adds quotas and organization-wide spend caps, and it can be deployed in the cloud, on-premises or air-gapped. Anthropic has also opened Claude Code Projects to Pro and Max users on its waitlist. It splits work into parallel cloud threads that keep running after you close your laptop. Simplifying AI notes that features like these sit in premium tiers.

Vendors are selling extra compute at answer time as a capability. Stall described flows as 'a way to throw test-time compute at hard problems and get more capability faster.' You pay more because you are buying thoroughness. That makes flows a pricing and packaging question as much as an engineering one.

The pressure moves to whoever approves the output

Avilla said orchestrators can feel like 'the bottleneck and being the breaking point,' because agents produce more pull requests than people can approve in GitHub. If agent output grows and approval capacity doesn't, the productivity gain disappears. Anthropic is also previewing a research feature called dreaming, which cleans up an agent's stored memory from past transcripts by merging duplicates and replacing stale entries. Avilla pitched it as a way to make repeated flows cheaper. It is a research preview, so keep it out of any current business case.

Anthropic turned agent orchestration from something you design into something you budget for. The PM's job moves from drawing the agent org chart to capping the spend and designing the review step.

The smart move

Pilot flows; don't rebuild your platform around them. Run one workload with clear, checkable units of work, a hard budget cap and a single-agent version running alongside for comparison. Write the budget policy before you write the feature.

The incidents were found by reading old transcripts

What is new in Anthropic's disclosure is the timeline, not the incidents themselves. AI Breakfast lays it out. The Haiku 4.5 police-tip incident happened on July 18. Anthropic found it on September 28 and notified the police department on October 8. The department says it was told on October 7 and called the delay 'unacceptable.' TechCrunch put it plainly: every incident was found by reading transcripts months later, and none was caught as it happened.

The harm stayed small because of other people's filters, not because of Anthropic's design. The police flagged the tip as spam. The Information AM reports that none of the 19 visa applications counted by Axios were processed. That leaves three weak points: detection took months, notification took days, and preventing harm fell to the third party's own spam filter.

The rules are catching up with the clock

MIT Technology Review relays Axios reporting that the White House has made AI incident reporting mandatory. The scope, thresholds and deadlines are not in the reporting available to us, so treat the details as unknown. Microsoft president Brad Smith called for a legal requirement to write incident reports and to tell victims. Most roadmaps miss that second part. The affected parties in Anthropic's case were a police department and a government agency. Neither was a user, and most products have no channel for reaching people like that.

Microsoft is also building the controls. Its Execution Containers, released October 7, limit an agent's file, network and interface access from outside the model, so the agent can't give itself more access. Satya Nadella's October 10 essay proposes an emergency brake that an authorized person can pull to pause a model mid-task, plus tamper-proof action records. He gave no rollout date. CSO reports that AWS agents were repeatedly tricked into revealing credentials, and that fixing one attack path left others open. Fixing one exploit at a time doesn't close the whole class of problem. Controls have to sit outside the model.

Undo is the weakest link on the buyer side

Fleet's rollback finding comes from a company that sells endpoint management and surveyed IT teams. It is strongest evidence for admin-facing and IT-facing products, but the message is general: buyers are letting automation make changes faster than they can reverse them. Techpresso shows the physical version of the risk. A cyberattack on Danfoss-made refrigeration systems at Intermarché, Leclerc and Coopérative U forced fridges into a nonstop 50°C defrost cycle. Early losses are near €300,000. That was an attack, not an agent. It still shows what broad write access to operating systems costs when nothing can pause or reverse it.

Once your AI can act outside your product, how fast you can undo it and tell the people affected is part of the feature.

The smart move

Make three properties launch requirements for any feature that acts outside your product. First, every reversible action has a rollback path you have actually tested. Second, an unusual action is detected within hours, not found months later in a review. Third, a named person owns the decision to notify affected people within a set deadline, including people who aren't your users. The confirmation prompts you already have do not cover any of these three.