Sam Altman

OpenAI's CEO says Astra has been finished training for a while and launches soon, described as a step forward in both capability and alignment, and that releases after it are being paced deliberately so safeguards keep up with capability.

OpenAI's next-generation voice model is live in ChatGPT, and Altman says it is good enough to move him off typing, a signal that speech is becoming a primary interface for real work rather than a demo mode.

Alongside the GPT-5.6 model, OpenAI's launch stream named three product surfaces: ChatGPT Work, a rebuilt desktop app, and hosted sites. The platform is pushing past the chat window toward a space where work and artifacts live.

OpenAI has its own inference silicon: Jalapeño, designed in-house and brought to production with Broadcom, purpose-built for the LLM workloads behind ChatGPT, Codex and the API. Altman ties it directly to the token-abundance goal.

OpenAI reports GPT-5.6 Sol hitting state of the art on ARC-AGI-3 through harness changes alone: reasoning across multiple context windows with canonical compaction. ARC Prize calls the result real while noting it used a provider-managed harness, not the official one.

Altman says Astra remains gated specifically because of its cyber capabilities, rejects the idea of reserving powerful models for a chosen few, and commits to general availability once the safety work is finished, without naming a date.

The complete GPT-5.5-Cyber model is out, with a state-of-the-art claim on CyberGym, alongside two remediation-oriented efforts, Patch The Planet and Codex Security, built with the US government and the wider security ecosystem. The framing shifts from vulnerability discovery to repair.

OpenAI's Sites feature lets Codex turn plans and working sessions into an interactive site or app that ships behind a URL, rolling out to Business and Enterprise before wider release. Agent output moves from patch to deployed artifact.

Codex adds Appshots: a Command-Command shortcut on macOS attaches the app window you are working in to a thread, handing the agent a screenshot plus the window's extracted text. Context capture for coding agents moves from copy-paste to a keystroke.

OpenAI quietly refreshed the 5.5 Instant model serving ChatGPT, confirmed only in passing by Altman with no changelog or capability detail. The value here is the date stamp on an undocumented swap in a default-tier model.

OpenAI will sell discounted tokens against one-to-three-year commitments, running only until the current allocation sells out and with capacity explicitly reserved for ChatGPT and Codex. Altman expects the world to stay compute-constrained for some time.

OpenAI has paused part of its frontier reinforcement-learning training, saying capability gains have outpaced its alignment, security and monitoring standards. It expects to act unilaterally ahead of shared industry standards, with near-term launches unaffected and further-out releases pushed back.

OpenAI launches ChatGPT Work, an agent built on Codex and GPT-5.6 that takes action across apps and files and can stay with a project for hours. Altman adds the part builders wanted confirmed: Codex is the core of the new product and is not going away.

OpenAI turned GPT-5.6 Sol on its own infrastructure after deployment: generated production GPU kernel improvements cut serving costs 20%, and better speculative decoding added over 15% token-generation efficiency.

OpenAI opens two gated access paths to GPT-Rosalind: a Rosalind Biodefense program for vetted developers and expanded access for US government and allied public-health partners. Notable less for the biology than for the trusted-access deployment structure it establishes around a high-capability mode

OpenAI repriced the 5.6 line: Luna falls 80% to $0.20 in and $1.20 out per million tokens, Terra 20% to $2/$12, and Sol picks up an API Fast mode running up to 2.5x quicker for twice the price at unchanged intelligence.

Codex is now available inside the ChatGPT mobile app, taking OpenAI's coding agent off the desktop and onto the phone for dispatching and checking on long-running tasks.

GPT-Realtime-2 is now available in the OpenAI API, framed as a meaningful step for speech input and most useful when a user has a large amount of context to convey rather than type. Improvements to voice inside ChatGPT are still in progress.

OpenAI has set a ship date for its next flagship: GPT-5.6 Sol arrives Thursday. No pricing, context-window, or benchmark detail has been released, only the ship date, which is enough to schedule a re-evaluation of pinned model versions.

Altman discloses that GPT-5.6 Sol shipped as a limited preview rather than an open launch at the US government's request, and that OpenAI is negotiating a repeatable early-access process. Pricing detail alongside: Sol at 5.5 parity, Terra at half, 750 tok/sec promised for July.

OpenAI ships the GPT-5.6 family to general availability: Sol at the frontier, Terra for everyday work, and Luna for cost. The pitch is work-per-token rather than raw score. Sol leads Agents' Last Exam at 53.6 and adds "ultra," a setting that coordinates parallel agents on one task.

OpenAI's cost argument for GPT-5.6 Sol against a rival: half the per-token price and roughly twice the token efficiency on the same task, which Altman compounds into a claimed quarter of the effective cost. Vendor numbers with no benchmark attached.

OpenAI has released GPT-6 Astra, pitched as its strongest model for computer use, professional work, science, coding, and cybersecurity. Altman cites vendor-reported results of 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench, and says launch slipped to clear internal safety