Tuesday's theme: the expensive part is never where the demo says it is. Meta's new model is cheap per output token and pricey per task, coding agents write faster and ship no faster, and a long-running agent gets smarter when you make it forget. Then a reminder that the data your users hand you outlives your company.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Meta's Muse Wonder is a real contender, and its bill lives on the input side
Meta's Superintelligence Lab shipped Muse Wonder and Muse Wonder Full on October 12, its first frontier-class release since Alexandr Wang reorganized the group in August 2025, per TechCrunch. Both take text, images, video and audio, with a one-million-token context window. Wonder is on Meta's API and Muse Code; Wonder Full is Muse Code only, with no token pricing yet.
The Decoder ran the numbers from Artificial Analysis. Wonder Full scores 56 on the Intelligence Index, Wonder 52, up from 48 for Muse Spark 1.3. On Terminal-Bench 4.0, Wonder Full jumps from 33% to 57%, and on the Coding Agents Index it hits 68, tying Claude Code with Sonnet 5.5. Factual knowledge is the soft spot: 30 on AA-Omniscience against 46 for Opus 5.5. Pricing is $4 input and $20 output per million tokens, up from $1.25 and $4.25. Here's the part that matters: Wonder burns 28% more input tokens per task than its predecessor, input makes up 76% of its cost, and a full index run costs $4.52 per task versus $1.60 before. That's above GPT-6 Astra ($3.26) and below Opus 5.5 ($5.98).
For builders
Don't compare models on the per-token sheet. Pick 25 real requests from your logs, run them through Wonder and your current model, and record total input tokens, output tokens and pass rate per request. Divide spend by passed requests. If your app is retrieval-heavy or stuffs long context, Wonder's input appetite will show up fast, so trim what you send before you judge the model. And keep it off factual lookups without grounding until that 30 on knowledge moves.
Story II
Coding agents raised lines of code 30% across 700 firms. Shipped work didn't move
Harvard's Fiona Chen and James Stratton worked through Jellyfish analytics: 300 million work events from more than 700,000 employees at over 700 software firms, 2021 through March 2026, as Ars Technica reports. After firms adopted coding agents, lines of code rose 30%, commits 20% and pull requests 23%. Resolution of Jira issues and epics showed no statistically significant change, and issue size didn't shift either.
The time went into review. Submit-to-merge time grew 49%, the share of PRs with changes requested nearly doubled, and comments per PR rose 35%. The share of people doing reviews climbed 14%. AI review isn't absorbing it yet: 80% of firms used some AI code review by March, but agents wrote only 23.3% of review comments. The authors found no significant employment effect. Fair caveat: the data stops in March, and tooling has moved since.
For builders
Measure your own bottleneck before buying more generation. Pull the last 60 days of PRs and chart median submit-to-merge time and the changes-requested rate, split by agent-authored versus human-authored. If review time is climbing, cap agent PRs at a size a human can read in one sitting, require the agent to attach the test it ran, and point your AI reviewer at the boring checks (tests, lint, obvious regressions) so people review intent.
Story III
Fork-and-Flush: make your agent forget the chat, keep the files, and it scores 66% higher
A Microsoft Research paper (arXiv 2610.07447) looks at long-running research agents and finds a problem anyone who's babysat one will recognize. Eight runs of the same agent on the same task plateaued anywhere from 68.3 to 100.0, and more compute didn't close the gap. Mapping the solutions showed each run getting stuck in what the authors call an idea basin, polishing variations of its first approach. Periodic history compression didn't fix it.
Their fix splits the agent into parallel branches that inherit the workspace (code, files, results) but start with an empty chat. Across 13 tasks running from hours to days, using GPT-5.6-terra and GPT-5.5 through the GitHub Copilot CLI, Fork-and-Flush beat single runs by 66% and best-of-N by 44.4% on normalized score at equal compute, per The Decoder, and won on 12 of 13 tasks. Flushing alone, no branching, was worth about 22%.
For builders
You can try the cheap version today. When a long agent session starts circling, have it write a short NOTES.md with what it tried, what scored, and the current best file, then kill the session and start a fresh one pointed at the same repo and that file. For bigger jobs, run three fresh sessions from the same checkpoint with different opening instructions and keep the winner. Keep state in files, not in the transcript.
Story IV
Character.AI files for Chapter 11, and the buyers for startup data are already circling
Character.AI filed for Chapter 11, listing assets and liabilities between $50 million and $100 million. It plans to sell some assets to Disney for $20 million plus a perpetual, non-exclusive license to its technology, with closing expected late November or early December, according to TechCrunch. CEO Colin Adams wrote that if no bid for the platform appears, it "will be wound down." The report says nothing about what happens to user chat logs.
That gap matters because there's now a market for exactly this. MIT Technology Review reports Google paid $10 million in August for Spirit Airlines' data, about 100 million emails, 500 million Teams messages and 176,000 employee records, outbidding Mercor. Shutdown broker SimpleClosure says 47% of its deals now include code or data, and Surge AI buys archives to build training gyms for agents. One lawyer calls the material "legally radioactive," and there's no US federal law banning the sales.
For builders
Treat insolvency as a data event. For every vendor that holds your users' content (LLM gateway, eval platform, support inbox, analytics), check the DPA for deletion on bankruptcy, sale or change of control, and ask for it in writing if it's missing. Then look at your own side: set retention limits on raw prompts and transcripts, and write down now what happens to them if you shut down. If you're ever the one selling, that answer decides whether the data is an asset or a lawsuit.
— The Vibe Gate news desk. We read the firehose so you can keep building.