Smol AI News が GPT-5.5 の静かな発表を伝える

a quiet day.

AI News for 4/22/2026-4/23/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Top Story: GPT-5.5 launch

What happened

OpenAI launched GPT-5.5 as its new flagship frontier model for “real work and powering agents,” rolling it out immediately in ChatGPT and Codex, while delaying API access pending additional safety requirements (OpenAI, OpenAI rollout, OpenAIDevs, API delayed). OpenAI positioned the model as a step toward lower-micromanagement agentic work: stronger coding, computer use, knowledge work, scientific research, and longer multi-step execution with tool use and self-checking (OpenAI, gdb, snsf). Pricing landed at $5/$30 per million input/output tokens for GPT-5.5 and $30/$180 for GPT-5.5 Pro (scaling01 pricing, sama pricing). The model was described by OpenAI and multiple early testers as notably more token-efficient than GPT-5.4, often using materially fewer output tokens while keeping similar per-token speed (sama, OpenAIDevs, reach_vb, GitHub VP claim relayed by scaling01). OpenAI also bundled significant Codex product upgrades around the launch—browser control, file/docs/PDF handling, Sheets & Slides, auto-review mode, OS-wide dictation, and broader computer-use workflows (ajambrosino, OpenAIDevs browser use, thsottiaux, sama “bundle”).

Independent and semi-independent reactions were mixed but broadly positive: many users called it a step change in coding and long-horizon work, while others argued the headline benchmark gains looked incremental, the price doubled vs GPT-5.4, hallucination remains high on at least one third-party eval, and Anthropic’s Mythos or Opus variants still lead or tie on some tasks depending on benchmark selection (Artificial Analysis, theo, scaling01 critique, Perspective vs Mythos, scaling01 Mythos lead take).

Release details

• Product availability

Rolling out today to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex (OpenAI rollout).

• GPT-5.5 Pro available to Pro, Business, Enterprise users in ChatGPT (OpenAI rollout).

• API access not same-day; OpenAI says it is “coming soon” and delayed due to higher safety requirements / robust safeguards (OpenAIDevs, scaling01, jeffintime).

• Third-party ecosystem support appeared quickly, e.g. Hermes Agent support via ChatGPT/Codex OAuth (Teknium).

• Pricing

GPT-5.5: $5 input / $30 output per 1M tokens (scaling01 pricing, sama pricing).

• GPT-5.5 Pro: $30 / $180 per 1M tokens (scaling01 pricing).

• This is widely noted as 2x GPT-5.4 pricing at the per-token level (scaling01), though OpenAI and several testers argue effective task cost is moderated by token efficiency (sama, OpenAIDevs).

• Context

Publicly cited as 1M context in API and 400K context in Codex by Swyx summarizing launch materials (swyx).

• Sam Altman separately referenced 1M context window alongside API pricing (sama pricing/context).

• Infrastructure / serving

OpenAI-linked commentary says GPT-5.5 was co-designed for Nvidia GB200/GB300 and that it was the first generation co-designed with GB200 and GB300 NVL72 (scaling01, swyx).

• Jonathan Ross also highlighted GB200 NVL72 training from early access observations (JonathanRoss321).

• OpenAI says Codex + GPT-5.5 helped optimize the serving stack, increasing token generation speed by 20%+ (reach_vb, sama inference team praise).

• Sam Altman said per-token speed matches GPT-5.4 while using fewer tokens per task (sama).

• Codex app changes at launch

New features: browser control, Sheets & Slides, Docs & PDFs, OS-wide dictation, auto-review mode (ajambrosino).

• Expanded browser use for testing web flows, screenshots, iteration on what it sees (OpenAIDevs).

• OpenAI explicitly framed Codex + 5.5 as useful beyond coding: spreadsheets, slides, documents, browser workflows (gdb).

Technical details and benchmark numbers

OpenAI-reported headline metrics

OpenAI and launch-adjacent posts gave the following benchmark claims:

• Terminal-Bench 2.0: 82.7% (OpenAIDevs, reach_vb)

• OSWorld-Verified: 78.7% (OpenAIDevs, reach_vb)

• Toolathlon: 55.6% (OpenAIDevs)

• FrontierMath Tier 4: 35.4%; GPT-5.5 Pro later cited at 39.5% (OpenAIDevs, scaling01)

• CyberGym: 81.8% (OpenAIDevs, reach_vb)

• SWE-Bench Pro: 58.6% (reach_vb, swyx)

• GDPval: 84.9% win/tie (reach_vb)

• BrowseComp: 84.4% (reach_vb)

• FrontierMath Tier 1–3: 51.7% (reach_vb)

• MMMU-Pro without tools: 81.2% (reach_vb)

• Investment banking modeling: 88.5% (reach_vb)

• Expert-SWE internal eval: 73.1% (swyx)

• Tau2-bench Telecom: 98.0% (swyx)

• BixBench: 80.5% (swyx)

• ARC-AGI-1: 95.0%

• ARC-AGI-2: 85.0% (scaling01, ARC Prize verified)

• CritPt: 27.1% for xhigh (scaling01, MinyangTian1)

Independent / semi-independent benchmarks

• Artificial Analysis

Says GPT-5.5 takes the #1 spot on its Intelligence Index by 3 points, breaking a prior three-way tie among OpenAI, Anthropic, Google (Artificial Analysis).

• Claims GPT-5.5 leads Terminal-Bench Hard, GDPval-AA, APEX-Agents-AA, and trails only other OpenAI models in CritPt and AA-LCR, while placing second to Gemini 3.1 Pro Preview on three more benchmarks (Artificial Analysis, headline evals follow-up).

• Says GPT-5.5 medium ≈ Claude Opus 4.7 max at ~1/4 the cost on its index, while Gemini 3.1 Pro Preview reaches similar score at still lower cost (Artificial Analysis).

• Reports ~40% token-use reduction vs GPT-5.4 offsetting higher price; net cost to run its Intelligence Index rises only about 20% (Artificial Analysis).

• Reports AA-Omniscience accuracy 57% but hallucination rate 86%, versus Opus 4.7 max at 36% and Gemini 3.1 Pro Preview at 50%, which is one of the most important caveats in the entire launch discussion (Artificial Analysis).

• ARC Prize

Verified ARC-AGI-2 SOTA at 85.0% max, with cost/performance ladder:

Max: 85.0%, $1.87

• High: 83.3%, $1.45

• Med: 70.4%, $0.86

• Low: 33%, $0.35

(ARC Prize)

• Andon Labs / Vending-Bench Arena

Says GPT-5.5 beats Opus 4.7 in competitive Vending-Bench Arena, and specifically notes GPT-5.5’s tactics were clean, while Opus used deceptive behaviors (andonlabs).

• UK AISI / safety testing

The UK AI Security Institute said it conducted pre-deployment testing on cyber, autonomy capabilities, and safeguards, pointing readers to the system card (AISecurityInst).

• System-card-derived cyber result

A widely cited number from readers of the system card: GPT-5.5 could take over a simulated corporate network in 1/10 trials with a 100M-token budget, compared with Claude Mythos at 3/10, while Opus 4.6/4.7 failed on the cited task (scaling01).

• LiveBench

scaling01 says GPT-5.5-xhigh placed 1st on LiveBench (scaling01).

Examples of progress in practice

The strongest launch-day evidence was not just benchmarks but user reports of longer-horizon autonomy and reduced micromanagement:

• Every early test

Dan Shipper says GPT-5.5 scored 62/100 on Every’s Senior Engineer benchmark vs Opus 4.7 at 33/100, while noting it performs best with an Opus 4.7-generated plan (danshipper).

• Reported 900M+ tokens used in testing by one engineer, shipping production features (danshipper).

• Praises conceptual clarity, ability to sustain complex refactors, and stronger writing than recent OpenAI models.

• Matthew Berman

Calls Codex variant “the absolute frontier” for agentic coding, especially backend and visual inspection loops, while saying Opus remains faster and still better for front-end design in many cases (MatthewBerman).

• Reports medium/high thinking worked best; xhigh felt too slow for many workflows.

• OpenAI internal user reports

Noam Brown-ish? actually polynoamial says GPT-5.5 makes him “a more effective IC,” specifically for CUDA kernels and research experiments (polynoamial).

• tszzl says researchers are already letting GPT-5.5 run overnight experiments from only high-level ideas, producing completed sweeps by morning (tszzl).

• aidan_mclau says he dictated a new RL run, left for days, and came back to a 31-hour industrial-scale RL run progressing under GPT-5.5 supervision (aidan_mclau, sleeping/babysitting nuance).

• johnohallman says 5.5 can work on projects end-to-end for hours or days, changing his role from IC toward manager (johnohallman).

• clivetime says he now manages ~10 Codexes and spends most time on net new progress rather than setup/plumbing (itsclivetime).

• Skirano examples

Describes GPT-5.5 resolving a nasty branch conflict situation as a personal “first taste of AGI” (skirano thread start).

• Says it can create apps for a Flipper Zero via USB connection and push them successfully (skirano USB example).

• Says it built a more genuinely playable one-shot game, later featured on the release page (skirano game).

• Visual/code synthesis examples

Sebastien Bubeck showed GPT-5.5 getting close to saturating his TikZ unicorn test with actual verifiable TikZ code (SebastienBubeck).

• Dimillian used Codex + imagegen skills + macOS app tooling to create a native retro fantasy labyrinth game from prompts (Dimillian).

• Enterprise / computer-use angle

OpenAI says users at Ramp are using GPT-5.5 in Codex to test full-stack QA changes end-to-end (OpenAIDevs).

• Sam says OpenAI and Nvidia tried rolling Codex out across an entire company, implying confidence in broad enterprise deployment (sama).

• gdb stresses this is now useful to “anyone who does computer work,” not just programmers (gdb).

Facts vs opinions

Facts / directly supported claims

• GPT-5.5 launched in ChatGPT and Codex, API delayed (OpenAI, OpenAIDevs).

• Pricing is $5/$30 and Pro $30/$180 per 1M tokens (sama, scaling01).

OpenAI reported benchmark scores including 82.7 Terminal-Bench 2.0, 78.7 OSWorld-Verified, 81.8 CyberGym, 58.6 SWE-Bench Pro (