No. 032 · 16 September 2026 · 3 min← No. 031 · No. 033 →

The one thing

01

TypeSafe prices Jev at $0.042 per million input tokens with free output, then gates it behind a waitlist.

Diogo Almeida, who worked on the instruction-following research behind ChatGPT, came out of two years of stealth yesterday with what TypeSafe AI calls a System One model. Jev takes unstructured program state in and returns typed probabilistic decisions with calibrated confidence; it generates no strings, so it cannot invent a field that was never declared. Published price is $0.042 per million input tokens with output free, against the $0.20 to $10 per million TypeSafe lists for frontier models.¹

The speed claim is 70ms to 500ms end to end, which TypeSafe puts at 40x to 200x faster on the query shapes it targets — from evaluations it designed, ran on its own laptops and scored against two frontier models it chose as the reference.¹ It took 1,085 points on Hacker News overnight.¹⁰ Access is a waitlist, and a price you cannot buy at is a press release.

VERDICT · TOO EARLY — Join the waitlist and keep the pricing table. Hold the verdict until somebody outside the company has run it on their own workload.

Shipped

02

Breaking

GitHub has disabled SHA-1 in HTTPS

On 15 September GitHub completed a sunset it had scheduled in advance: SHA-1 in HTTPS is now off for github.com and partner CDNs, including Enterprise Cloud and Enterprise Cloud with Data Residency. Enterprise Server is unaffected.²

MATTERS TO · Anyone whose pipeline still builds on an old base image. A TLS stack that only offers SHA-1 signature algorithms can no longer reach github.com, and it surfaces as a handshake failure in a clone that worked on Monday. Audit the oldest image in your CI first.

nofeed.dev/issues/2026-09-16/github-sha1-https-sunset/

Use it

Gemini 3.8 Live and 3.8 Live Extended Thinking, in the API today

Both are available now through the Gemini API and Google AI Studio at $0.005 per minute of audio input and $0.018 per minute of audio output.³ For a voice agent the useful part is asynchronous function calling — tool calls run in the background while audio keeps streaming — alongside visual grounding, mid-conversation switching across 97 languages, and alphanumeric parsing for confirmation codes.³⁴

MATTERS TO · If you have been stitching a cascaded stack out of a recogniser, a model and a voice, this collapses three hops into one, with Live API integrations through LiveKit, Pipecat and Vercel.[^3] Enterprise access is still private preview.

nofeed.dev/issues/2026-09-16/gemini-3-8-live/

Promised, not shipped
TypeSafe Jev — early access off a waitlist, no general availability date.[^1] · Gemini Enterprise for Customer Experience — "coming soon", no date.[^4]

The conversation

01

The claim — including ours

Three vendors published a leaderboard position yesterday. A preprint filed the same afternoon argues the leading coding-agent leaderboard can no longer order its own top entries.⁶

From the floor

The authors of arXiv 2609.17394Preprint read directly; no peer review
Audited 254 SWE-bench submissions across four splits without running a single model. On Verified the top two entries each resolve 396 of 500 instances, and the top ten share 285 successes and 51 failures — leaving 164 instances that separate them at all. Exact paired McNemar tests separate none of the 29 adjacent Verified top-thirty pairs at alpha 0.05, and within-model scaffold ranges reach 29.8 percentage points against the 8.8-point spread across that whole top thirty.⁶

Our take

On SWE-bench Verified the harness you wrap a model in moves the resolution rate by up to 29.8 points, more than three times the spread across the entire top thirty. A rank won by swapping scaffolds is a fact about your plumbing — so pick your agent on your own repository.

The X and Reddit layers did not cover this window: the newest digest for either is 2026-09-15T05:30Z, roughly 24.5 hours old, closing before ours opens, and the accounts roster produced only markdown. Today's community reading is Hacker News and arXiv alone.

In production

01

1.4 hours

Claude Mythos 5.1 and Fable 5.1 returned intermittent error spikes on 15 September, an incident Anthropic rated major impact and closed after one hour and twenty-four minutes.⁵ OpenAI ran a separate major-impact incident on gpt-image-2.5-flare for 0.7 hours.⁹Source · Anthropic and OpenAI status pages, read 16 September 2026

Durations computed from the status timelines, not quoted by either vendor.

Skip this

04

01

TypeSafe's workflow evaluations The company designed the evaluation, chose the reference models and ran it from its own laptops. We quoted its price and refused its benchmark in the same breath.[^1]

02

"#1 on Artificial Analysis Speech-to-Speech" A vendor citing a leaderboard on the morning it ships, while its own coverage puts the sibling model second elsewhere.[^8]

03

Strix's admin access to Baseten's GitHub 258 points on Hacker News yesterday, but dated 1 September and the token was rotated in July. Out of window, already fixed.[^7]

04

Twelve releases across the watchlist Claude Code 2.1.273, openai-python 3.14.1, gemini-cli 0.60.0, adk-python 2.9.1, anthropic-sdk-python 1.6.0 and seven more, every one with an empty release note.

Everything we saw

141

141 candidates scanned · 3 used in this issue — the rest, with the reason each one was left out

The receipt. Ranking is only trustworthy if the discarded pile is visible, so here it is: everything the collectors surfaced in the window, with its signal and what we did with it.

Item	Source	Signal	Call
TypeSafe AI: System One models and Jev	Vendor blog / HN 1,085p	Priced, waitlisted, vendor-run evals	Led the issue as TOO_EARLY
SHA-1 in HTTPS on GitHub sunset	GitHub changelog	Completed, breaks old TLS clients	Shipped, BREAKING
Gemini 3.8 Live and 3.8 Live Extended Thinking	Google developer blog	In the API today, per-minute pricing	Shipped, USE_IT
Coding Agents Have Converged (arXiv 2609.17394)	arXiv cs.SE	254 submissions audited, no ranking survives	Framed the conversation
Anthropic: error spikes on Mythos 5.1 and Fable 5.1	Anthropic status	Major impact, 1.4 hours	Printed as the production figure
OpenAI: elevated errors on gpt-image-2.5-flare	OpenAI status	Major impact, 0.7 hours	Printed alongside
Strix: admin access to Baseten's GitHub	Vendor blog / HN 258p	Real finding, remediated in July	Skipped, out of window
Enforce GitHub Advanced Security configurations	GitHub changelog	Enterprise admin scope only	No decision for a developer
Copilot suggests custom property definitions	GitHub changelog	Public preview, org metadata	Too narrow to print
Vercel: Is Agentic tailors its audit by site type	Vercel changelog	Report categorisation change	Watching
Hugging Face / IBM: will your agent do it again?	Hugging Face blog	Consistency tooling, no release	Watching
Java 27 and Swift 6.4	HN 325p / 122p	Large releases, off this beat	Off beat
AgentGuard: guardrails from anomalous agent traces	arXiv cs.SE	No implementation to run	Watching
GitHub trending sweep	Discovery layer	50 on-beat repositories, 2 admitted after judgement	Watchlist now 35
X accounts roster and keyword digests	Community layer	Newest collection 2026-09-15T05:30Z	Layer did not cover the window
Reddit practitioner digest	Community layer	Newest collection 2026-09-15T05:30Z	Layer did not cover the window

EOF

End of feed. That is everything from the window worth your time.
Next issue tomorrow, 06:00 UTC — and if nothing ships, it will say so in two hundred words.

Still developingClaude Code's weekly limit at 125% of the old standard, promised for 14 September, is still on no vendor page we can find — carried from issue 30. The pin has moved to 2.1.273. Copilot seat charging and the model cutover land 1 and 2 October; Cursor's bundled OpenAI access is proposed to end 12 November.