Ephemeris.№ 123 · ten stories from the week
03SaturdayOct ’26
Two cheaper models, a CLI written for agents to drive, and a 136-billion-token bill for a port to Rust.
The cheaper Claude now beats Opus at the terminal.
Anthropic released Claude Sonnet 5.5 on 28 September at Sonnet 5’s price, $2 in and $10 out per million tokens. It scores 70.6% on Terminal-Bench 4.0, ahead of Opus 5.5’s 66.4%, generates more than 30% faster than Sonnet 5 and costs up to 30% less per task. It is the first Sonnet to ship with cyber safeguards. If you run Sonnet with thinking off, switch to the new between_tools setting when you move.
70.6%
Terminal-Bench 4.0. Sonnet 5 scored 10.3% on the same test.
- 55.5%
- CursorBench 4.0, second only to Opus 5.5
- 80.1%
- OSWorld 2.1, the computer-use benchmark
- $0.20
- per million cached input tokens; model id
claude-sonnet-5-5
Near-Astra, for a fifth of the price.
OpenAI made more than 20 announcements at DevDay on 29 September. Start with GPT-6.1 Sol: it matches GPT-6 Astra on DeepSWE v1.1 at roughly a fifth of the cost, and the API lists it at $2 in, $10 out and $0.10 for cached input per million tokens. It runs in the API as gpt-6.1-sol, in ChatGPT Work and in Codex, though not yet in Chat.
- 1GPT-6.1 Sol
2.2 points above Opus 5.5 on AutomationBench at medium effort, for about a third of the cost.
- 2Ultrafast
A paid speed tier: up to 8× faster generation in Codex (300 tokens a second), up to 6× in the API. Astra first, on Pro 500 and Enterprise.
- 3Dots
Always-on agents on GPT-6 Astra, each with its own cloud computer and 4,000+ app connections. Pro and Business Premium.
- 4MCP Events
Plugins can start an automation when something changes in a connected app, through the proposed MCP Events spec. All plans.
One developer, 128 pull requests, 832,000 lines of Rust.
Stephen Toub moved the GitHub Copilot agent runtime, the engine under the Copilot CLI, the Copilot app and the latest VS Code, from TypeScript to Rust between 12 May and 21 August, with agents writing most of it. Each pull request replaced one component with Rust behind a shim and deleted the TypeScript, so the port shipped as it went. The bill was about 136.3 billion tokens, roughly $120,000 at a 96.22% prompt-cache hit rate, plus about three weeks of his own time.
The numbers
- 5.25 s → 292 ms
client, session and one turn, TypeScript against in-process Rust - 7.55 → 120
session lifecycles a second under 100 concurrent pipelines - 1,383 MB → 126 MB
peak memory added by a ten-client batch
The rule
“Protect the oracle from the agent.”
Every end-to-end test ran against the Rust code at each step. The lesson Toub draws: an agent changing an implementation must not also get to redefine correctness by weakening a test or updating a snapshot.
The habit
The agents did ten times as much reading and searching as editing. Toub’s first takeaway: “the agents spent far more time gathering evidence than changing code.”
$ Agents ran 48% of Wrangler. Now they get their own CLI.
Agents accounted for a quarter of Wrangler use in March and 48% last week, and Wrangler covers about 280 of Cloudflare’s operations. cf is generated from Cloudflare’s OpenAPI schemas and covers about 3,000. Output is JSON by default, cf cli search finds a command from a plain-language request, and config moves to a type-checked cloudflare.config.ts, with Vite as the dev server. Open beta since 28 September.
$ npm i -g cf # convert a Wrangler project to cloudflare.config.ts $ cf migrate # describe the job in plain words, get the command $ cf cli search "…" # JSON out by default, compact for agents $ cf d1 list
74.4% at 4.6¢ a ticket became 98.9% at 1¢.
Anthropic’s claude-api skill for Claude Code got two commands on 28 September. /claude-api build-eval interviews you, builds an eval in your repo from production transcripts and tickets before it writes synthetic cases, and runs the grader twice on the same output to catch flipped verdicts. /claude-api hillclimb splits the cases into train and test, makes one change per round, and flags a change that lifts train while test stays flat. Run claude update first.
| Round | Model, effort | Accuracy | ¢ / ticket |
|---|---|---|---|
| Baseline | Opus 4.8, high | accuracy74.4% | ¢ / ticket4.6 |
| Round 1 | Opus 5.5, low | accuracy87.8% | ¢ / ticket1.9 |
| Round 2 | Sonnet 5, low | accuracy88.9% | ¢ / ticket1.0 |
| Final | Sonnet 5, low, prompt fixed | accuracy98.9% | ¢ / ticket≈1.0 |
Anthropic’s customer-support benchmark. On the 14 held-out tickets the search never saw, the final setup scored 90.5% against the original’s 78.6%, at about a fifth of the cost.
Bayram Annakov ran hillclimb on lead scoring at Onsa: a 616-case run for $17 cut cost 25% and ran 2.6× faster, while lost good leads rose from 1.4% to 4.3%. It failed the test. He shipped it anyway and scheduled a check for a week later.
63 million agent calls. A quarter were raw SQL.
PostHog counted 63 million tool calls to its MCP server from 130,000 people in the 90 days to 16 September. Agents mostly skip the purpose-built tools: raw SQL is 25% of calls, schema reads another 7%, and the SQL is where they fail, 4.4% of the time. Among third-party clients, Claude Code sent 32% of calls, Codex 22% and Cursor 10%. To track your own MCP server the same way: npx @posthog/wizard mcp-analytics.
7% of calls came from a chat window. The rest came from editors and terminals.
- 252,000
- insights created by agents in 30 days, to 99,000 by humans
- 2% / 5.7%
- error rates, Codex against Claude Code, on the same server
- 144,000
- tokens at p95 for one call that reorders dashboard tiles
“That’s an entire context window just to move a tile.”
Mobile has no revert.
A bad web deploy can be reverted in minutes; a bug shipped inside an app binary has to travel the whole path out again, through a fix, testing, store submission and review. So before the holidays Etsy’s mobile team narrows what can ship and what can page. Two alerts are allowed to be noisy: crash-free rate, and crashes per hour outside their historical norm. Everything else goes to a triage channel, and experiment flags switch off a bad variant without a release.
- IMid-October. Merge mindfully: nothing that makes a seller learn a new workflow while they stage holiday stock.
- IIThen. Nothing that taxes infrastructure; any experiment that adds load goes through the scalability team.
- IIIThanksgiving to Cyber Week. Nothing ships without very senior sign-off, and only to make things better.
“Week one, people read the notifications. Week two, they skim. Week three, they write a filter.”
Crash cost, reported to leadership monthly: sessions with a crash × conversion rate × average order value = estimated sales lost.
The lead plans. The sidekick does the work.
Devin Fusion, in Cognition’s Devin Desktop and CLI since 11 September, runs two agents at once: a frontier lead that plans and reviews, and a cheaper sidekick, SWE-2, that executes, runs the tests and reports back. They exchange briefs, results and feedback rather than whole transcripts, so each keeps its own prompt cache. Cognition’s conclusion is to judge models by price per task, since a model that costs more per token can still finish the job for less.
Lead
Fable 5.1
Owns the plan, interprets ambiguity, reviews.
Sidekick
SWE-2
Implements the change, runs tests, reports back.
64.3 · $14.63DeepSWE 1.1, Fable 5.1 alone, per task
63.1 · $7.88Fable 5.1 leading SWE-2, 46% cheaper
“Fable delegated earlier and gave better briefs, while Opus micromanaged the sidekick and redid much of its work.”
Try it: curl -fsSL https://cli.devin.ai/install.sh | bash
Two YAML files found 24 Android bugs.
GitHub Security Lab gave its open-source Taskflow Agent two Android workflows. One sorts an app’s entry points into mobile and non-mobile; the other checks each one against a fixed list of bug classes, such as confused deputies and insecure broadcasts. Pointed at open-source apps, they found 24 vulnerabilities. You can run the same audit on your own repository.
OsmAnd · 10M+ downloads
An exported map activity trusts the extras in any incoming intent. An app with no permissions can import new settings, point the map tiles at its own server and read off the user’s location and every route’s start and end.
Wikipedia for Android
A loose hostname check on a deep link and a loose cookie-domain check chain together. One tapped link hands over a session token that works across every Wikimedia project.
./scripts/audit/run_mobile.sh myorg/myrepo
From a codespace on seclab-taskflows, with a Copilot licence. An hour or two on a medium-sized repo. GitHub’s own caveat: models find the bugs and misjudge their severity, so a researcher reviews each one.
Patch to 16.3.8 or 15.5.27.
Next.js 16.3.8 and 15.5.27, out 30 September, fix seven vulnerabilities. The one rated high is an SSRF in Image Optimization, and you are exposed only if you set images.remotePatterns. On 22 September, 16.3.6 fixed a critical remote-code-execution bug in the Node.js ImageResponse, inherited from the Satori library; the Edge version was unaffected.
SSRF to private IP ranges through an allow-listed remote image URL.
Cache poisoning of SSG and ISR pages, root-param and Draft Mode leaks through 'use cache', and metadata images that ignore dynamicParams.
The next dev MCP endpoint doesn’t check origin, so a malicious site you visit can read source snippets and logs.
# on 16.3 npm install next@16.3.8 # on 15.5 npm install next@15.5.27
That’s the week.
Issue 123, Saturday 3 October 2026. Ten picks from eight publishers, published between 11 and 30 September, eight of them since last Saturday’s issue. One came to us through a Telegram channel and one through a newsletter.
Rubric, in order: AI tools you could adopt this week · creative software and generative media · dev tools and agentic coding · privacy and security where the signal is high · science and research with a practical kernel · anything immediately actionable for a senior engineer or founder.
Also read this week and left out: The Zvi, Import AI, jeffbullas.com, Vercel, fly.io, @seeallochnaya, @TochkiNadAI, @denissexy, @rvnikita_blog. Anthropic Engineering published nothing new inside the window.