Quick Hits

• OpenAI's DevDay keynote got the usual liveblog from @simonw, which @HamelHusain calls "the most time-efficient way to get up to speed on what's launching today." In the same window, @philipkiely reported that enterprise teams can run open models like GLM-5.3 Flash and Kimi K3 natively in Codex, with spend counting against their OpenAI commit; @thsottiaux cheered it with "Open is the way."
• Cheap, fast decision models keep gaining ground: @OpenAIDevs announced a Decisions API powered by GPT-6 Luna (limited preview), and @muratcan described Jev forecasting outcomes across 2,029 real phone calls for about $3, which @natebjones reads as early evidence of what System One-style models can do.
• Constraints shaped two deployment stories: @alex_prompter amplified @mardehaym's thread on a HIPAA-compliant voice agent for healthcare billing calls, and @abliteration_ai relayed @eddylazzarin turning to unrestricted models for defensive security tasks after restricted ones refused.
• Two big claims, thin detail: @jrysana teased Rysana V2 as "up to 10,000x faster and 1000x more efficient," and @malikwas1f passed along @teortaxesTex's claim that DeepSeek's kernel work makes the Ascend ecosystem "fully viable for training."
• Elsewhere: @elonmusk postponed the Roadster demo to October 15 over high winds, @naval recirculated Musk's joke that a Delta flight is the best AI sandbox, and @leerob asked for feedback and feature requests for @Bot.

Open Models in Codex, With the Invoice Still Pointing at OpenAI

@philipkiely's one-liner is the day's most concrete enterprise item: open models like GLM-5.3 Flash and Kimi K3 can now be used natively in Codex, and the spend counts against an existing OpenAI commitment. @thsottiaux's reaction, "Proud of this. Open is the way," frames it as openness winning. The skeptic's read is that openness now arrives with a lock-in flavored billing arrangement: bring your own model, but the money still flows through the same commit. Both takes are opinions, and the underlying claim comes from a single post worth confirming against official docs.

Meanwhile @simonw liveblogged the DevDay keynote in San Francisco, a ritual @HamelHusain endorses as the fastest summary of what actually launched. If the announcements above landed at or around DevDay, no post in the feed says so explicitly, so treat the timing as suggestive rather than confirmed.

Decision Models Graduate From Hack to Product

@OpenAIDevs announced Decisions API, powered by GPT-6 Luna: define questions and possible answers to classify content, route requests, or choose an agent's next action. It is in limited preview. @sydneyrunkle's take is that "jev was the first, but certainly not the last," because decision models are cheap and low-latency enough to plug into model routing, guardrails, and tool or skill selection.

The most striking evidence comes from @muratcan's experiment, surfaced by @natebjones. Jev received 2,029 real phone calls from an AI receptionist, reduced to pure structure: turns, tool calls, workflow stages, and timing, with no audio or transcripts. It made 38,012 turn-level forecasts at 118 ms median latency, zero-shot with no fine-tuning, and by the halfway point separated calls that would book from those that wouldn't at AUC 0.78, ranking them correctly 94% of the time near the end. Total cost: roughly $3. These are author-reported numbers, and @muratcan concedes Jev over-focused on visible errors the agent usually overcomes. @natebjones stays measured: "we're very early on figuring out the implications of System One type models."

Tier Your Models: Opus Plans, Sonnet Swarms, Fable Advises

The best how-to of the day is @mirku21's Claude Code setup: run Opus 5.5 as the main session at high effort, hand explorer, worker, and researcher subagents to Sonnet 5.5 at medium effort, and keep Fable 5.1 on call as an advisor that only speaks up before a big plan, when an error repeats, and before calling a task done. One layer down, "Jev engineering" routes the forks that need no thinker, like which file or retry versus stop, to Jev in under half a second. The post even includes a paste-ready prompt that audits your config and shows diffs before editing anything.

The same size-matching logic appears as complaint and stunt. @zephyr_z9 says Astra is "too slow" and needs a smaller, faster model integrated, quoting @tenobrus's half-joke that OpenAI staff live in Slack and never feel the app's latency. And @charles_maddock ran GPT 6.1 Sol on a repo and offered $1,000 to the first engineer who guesses what the resulting PR does, which works as both a stunt and a quiet benchmark of how hard frontier-generated diffs have become to read.

The Best Agent Skills Come From People Who Shipped Without Them

@mattpocockuk named his top three skill makers: @poteto, @dexhorthy, and @emilkowalski. @Tao100086's breakdown shows the throughline: each spent years practicing a craft before encoding that judgment into agent-executable skills. Lauren Tan (@poteto), a React core member who has worked at Netflix, Meta, and Cursor, ships pstack, engineering skills that push agents through rigorous process, review, and verification before parallelizing work. Dex Horthy (@dexhorthy), HumanLayer CEO and author of 12-Factor Agents, focuses on context management and planning so LLM-based software survives real users. Emil Kowalski (@emilkowalski), a design engineer at Linear, writes skills covering animation curves, durations, and interface polish.

The market signal is @joesadoski, who uninstalled most Pocock skills and installed all of @poteto's, arguing the former now feel "too defensive against short horizon models" while pstack matches what the current frontier can handle. Skills, like models, apparently need re-evaluation as capability rises.

MCP Stretches From ChatGPT Extensions to Passport Offices

@mxstbr announced plugin extensions for ChatGPT: the same platform used to build features like meetings, health, and finance, built on MCP and MCP Apps. @RhysSullivan's praise is specific: it is "a great use of the extensions spec," notable because it composes with what developers already know rather than introducing a new walled surface.

@dillon_mulroy pushed the idea into policy: "every government service should be required to have an mcp." The trigger was @adambhaloo's announcement of a redesigned U.S. passport, with a new eagle motif, keepsake packaging, and a 2028 arrival. That announcement says nothing about MCP; the connection is purely @dillon_mulroy's argument that public services should be agent-accessible by default, not just screen-friendly.

Practical Takeaway

If you operate any agent harness, the strongest actionable thread today is explicit model tiering. Map your workflow for decision points (routing, guardrails, tool selection, pre-plan and pre-done checks) and test a cheap, fast classifier there before reaching for a frontier model, the way @mirku21 and @sydneyrunkle sketch, then measure cost and latency against accuracy, since @muratcan's $3 experiment only counts if the quality holds. And when adopting agent skills, weight the author's domain track record over prompt count, then revisit those skills as frontier models improve, per @joesadoski's uninstall.