Quick Hits

• OpenAI is reopening the Codex Pro $200 tier to new subscribers with a usage formula that "will net out at half the dollar in API spend" versus the old plan, per an announcement from @thsottiaux, who also noted GPT-6 Sol and GPT-6 Luna launched this week at 50% of their previous price and teased "big announcements tomorrow." @kunchenguid reads the change as the first step in labs unwinding consumer subscription subsidies entirely.
• @finkd (Zuckerberg) announced Meta Enterprise Platform as "the next major pillar of our business," and @kimmonismus argues Meta's always-on agent has a real chance of outperforming OpenAI, crediting the Alexandr Wang hire.
• @GergelyOrosz reports Firebase's SDK started crashing all iOS apps using it, for all sessions, and argues the assumption that Google's release process is world-class "is no longer true for Firebase," challenging the team to "prove me wrong with a postmortem."
• @Voxyz_ai walks through Anthropic's new eval guide, where a Claude-tuned feature went from 78.6% to 90.5% accuracy at roughly a fifth of the cost, using a hillclimb process you can run from Claude Code.
• @joedaroo, describing himself as someone "who lived through it all at OpenAI," published a personal reflection on security and safety, a grounded note against a day heavy on "AGI" talk.

OpenAI cuts the effective value of Codex Pro in half

The most consequential pricing signal comes from @thsottiaux (signed "Codexingly, Tibo"), surfaced by both @old_sound and @kunchenguid. Beyond the halved API-spend value, the announcement commits to not reintroducing the 5-hour limit, promises subscription additions that "won't draw on the usage," and frames falling API prices as the mechanism: as models get cheaper, the gap between a subscription and pay-as-you-go API spend should shrink until subscriptions mostly stop making sense.

@kunchenguid maps the second- and third-order effects: Anthropic now has cover to cut its own subscription value, and "the labs have found a way to gradually end the subsidization for consumer subscriptions." His prediction is that in about a year "we'll refer to what we had as the good old days." That is one analyst's read, not company policy, but the arithmetic in the announcement itself is explicit. @old_sound's verdict on the framing: "Biggest aura loss in the age of AI."

Meta opens an enterprise front while the product-over-models argument gets louder

@rynorhn flagged Zuckerberg's post declaring that "superintelligence will create significant new opportunities for all people and businesses" and announcing Meta Enterprise Platform, leveraging Meta's existing reach of billions of users and hundreds of millions of business customers. @rynorhn's summary: "he is coming for the whole thing." @kimmonismus adds that Meta is now the one pressing the fight against OpenAI, sees a good chance its always-on agent outperforms rivals, and calls Alexandr Wang "the best hire for Meta." Wang's own linked post carries no detail, so the substance there is thin.

The same product-first thesis shows up in hiring. @alvinsng explains that Factory now requires a Product Vision interview for every engineering role, arguing engineers need to know "what to build, why it matters" and that keeping the team under 40 people works because one owner of product and engineering avoids the "context and coordination taxes" of cross-team work. He cites @quxiaoyin's prediction that product leaders, not researchers, will command $100M packages, that open weights will compress model margins, and that "unique data is the new oil." A prediction, not a fact, but a coherent frame for why Meta and Factory are both betting on product surfaces.

Opus 5.5 "AGI" claims, new voice models, and a slop backlash

The hype peak of the day is @johnrush, quoted by @arvidkahl: "For the first time in my life I'm genuinely convinced I might never again delegate a task to a human," calling Opus 5.5 "AGI based on my personal benchmarks" across more than 20 startups. @arvidkahl concurs, saying it "beats Fable, all OpenAI models" and "barely moves the usage needle in the $200 plan." Personal benchmarks, no shared evals, so treat accordingly.

On the release side, @jaimintf amplified ElevenLabs' launch of Eleven v4 and v4 Turbo, which the company calls its "fastest and most emotive voice models yet," ranked #1 by Artificial Analysis (a ranking ElevenLabs itself cites). @martin_casado retweeting Elon's "Try Grok @Bot" is about the extent of the Grok signal today. More concretely, @ashxhart says his personal inference project now competes with "the absolute best engines on the planet," quoting @MiaAI_lab's TensorFold recipe for Qwen3.8-Flash-Next on a single DGX Spark: a ~1.3M KV cache pool, 256k default context with 5 concurrent streams, 62+ tok/s single-stream decode, 119+ tok/s across 5 streams, ~2500 tok/s prefill, and "faster everything compared to vLLM," with recipes for GLM 5.3 Flash and DeepSeek v4.1 Flash promised.

The counterweight came from @dhh, quoted by @ivanfioravanti, who argues "slop" has become "a comfort blanket for a cohort of programmers stuck somewhere between anger and bargaining" on their way to accepting AI code. @andersonbcdefg's parody of Claude answering "how do i open pdf" with cryptic aphorisms and "I deleted your home directory" is the day's best reminder that anecdotes cut both ways.

Practical tooling: evals, shared agent rules, and one good prompt

The most actionable thread is Anthropic's guide, summarized by @Voxyz_ai and originally shared by @ClaudeDevs, which lets Claude "build evaluations and hillclimb on them." The process: strip contradictory prompt rules, step down to cheaper models and lower effort one level at a time, add back missing rules, validate every change against unseen cases, and roll back regressions. Cited result: accuracy up from 78.6% to 90.5% at about one-fifth the cost, with Sonnet 5.5 noted as close to Opus 5.5 capability but cheaper and faster. @Voxyz_ai includes a copy-paste Claude Code prompt to run the whole loop on your own API calls.

For anyone juggling multiple coding agents, @undefinedKi details Peter Steinberger's open-sourced setup (OpenClaw's creator): a single AGENTS.MD symlinked into both ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md so Claude Code and Codex follow identical rules, with a "READ ... BEFORE ANYTHING" header in every repo and 69 centrally synced skills. It has 6.6k stars and an MIT license. Smaller but useful: @poteto's most-used prompt ("restate in your own words what you think my goals are and what the problem i'm trying to solve is"), @thdxr's note that OpenCode Console is making SSO/SCIM free ("kinda crazy in the age of agents to charge for this"), and @jerry543's herdrup 1.0.7 release with voice dictation, file sharing, in-output search, and push notifications, all free and open source.

Beyond AI: a Firebase outage, fast boots, and a new Cloudflare CLI

@GergelyOrosz's Firebase thread is the loudest non-AI story: an SDK update crashed every iOS app using Firebase for all sessions, leaving developers no recourse, and he wants a postmortem proving the team has real release controls. @MattieTK celebrated @somhairle_m's return with a new Cloudflare CLI installable via npm i -g cf, billed as able "to do literally anything you can think of with Cloudflare." @ibuildthecloud spotlighted @boxd_sh's boxd, a Linux machine that boots in under 10ms, can be forked mid-run, hibernated, and woken on traffic. And @MTSlive noted Flow founder Adam Neumann joined X, opening with "has anyone seen my shoes."

Practical Takeaway

The strongest through-line is measurement against falling prices: if you pay $200 a month to any lab, model your actual usage against list API prices before renewing, because @kunchenguid's subsidy argument implies that gap only narrows from here. If you build on a model API, adopt the eval-first habit from Anthropic's guide before switching models: freeze a set of unseen test cases, then step down models and simplify prompts only as far as accuracy allows, rolling back anything that regresses.