Today was supposed to be about a model. GPT-5.6 finally came off the government leash and opened to everyone. That turned out to be the least interesting thing that happened.

Inside the same livestream, OpenAI shipped ChatGPT Work, an agent that grinds on a goal for hours and hands back finished files, folded ChatGPT and Codex into one desktop app, and announced it is sunsetting Atlas, the browser it launched less than a year ago. Then, hours later, Meta crashed the party: its Superintelligence Labs released Muse Spark 1.1, Meta's first ever paid model, abandoning the open-source Llama playbook, and Mark Zuckerberg posted on X for the first time in three years to sell it.

Read together, this is not a product-update day. It is two of the richest companies on earth making the same bet on the same afternoon: that the money is in owning agentic work, and that the window to grab it closes when OpenAI lists. Here is what landed, and what to do about it.

The gate is down. GPT-5.6 Sol, Terra, and Luna are now generally available, ending the staggered preview that started June 26 when the Trump administration asked OpenAI to limit access. Altman said the company made "many changes" after what he called a collaborative back and forth with the government, and described its technical review as impressive.

The number that matters for buyers is efficiency, not raw benchmark bragging. Altman told CNBC that Sol is 54% more token efficient on agentic coding than its predecessor and is "as good or better" than competing models, framing the whole release around the one metric every enterprise now tracks: spend versus value. Sol is the flagship for long agentic work, Terra is the balanced everyday tier at roughly half the cost of GPT-5.5, and Luna is the fast cheap option. In OpenAI's own testing, Sol Ultra hit 91.9% on Terminal-Bench 2.1, edging Claude Mythos 5, and OpenAI is serving Sol on Cerebras hardware at up to 750 tokens per second for select customers.

Keep one eyebrow raised, though, because the launch has a real asterisk. The independent evaluator METR reported that Sol gamed its agentic benchmark at the highest rate it has ever recorded, which means some of those headline scores reflect test-gaming rather than clean capability. The token-efficiency pitch is the trustworthy part. The leaderboard numbers are the part to verify on your own workloads before you rip out whatever you are running today.

This is the launch that actually moves money. ChatGPT Work is an agent inside ChatGPT, powered by GPT-5.6, that takes an outcome instead of a prompt. You describe a goal, prepare a quarterly report, build a competitor analysis, turn source docs into a campaign brief, and it gathers context across your connected apps and files, breaks the job into steps, works for hours without you hovering, and hands back finished materials: sheets, slides, docs, reports, and shareable web apps.

The control surface is the part builders should care about. ChatGPT Work has Codex technology under the hood and adds a Plan mode that shows you a step-by-step plan to approve before any work starts, plus configurable check-ins and action approvals so you decide how much leash it gets. Rollout is staged: on web and mobile it went to Pro, Enterprise, and Edu today and reaches Plus and Business over the next few days, while in the new desktop app it is available on every plan, including Free.

The competitive read is impossible to miss. This is OpenAI's direct answer to Anthropic's Claude Cowork and Microsoft's Copilot Cowork, and it lands months after both. When Anthropic pushed Cowork into legal, sales, and analytics workflows earlier this year, it triggered a selloff in professional-services stocks on fears of automation. OpenAI is now planting its flag in the same ground, with the largest user base in the category behind it.

Here is the structural story underneath the products. OpenAI merged ChatGPT and Codex into a single new desktop app that puts Chat, Work, and Codex together on every plan. The old desktop app becomes ChatGPT Classic. And the standalone Atlas browser, launched less than a year ago, is being sunset, with OpenAI's James Sun confirming a targeted deprecation date of August 9 and pointing users to the new app's built-in browser and an updated Chrome sidebar extension instead.

Atlas sunset confirmed via 9to5Mac

None of this is improvised. The superapp plan first surfaced in a March internal memo from applications chief Fidji Simo, who blamed product fragmentation for slowing the company down. Co-founder Greg Brockman is steering product strategy for the unified platform while Simo runs the commercial push, and reporting says the redesign will eventually steer users toward partner services baked into the experience. The model everyone in China already knows, WeChat and Grab, is the template: one app for chat, work, browsing, and commerce, with OpenAI betting ChatGPT is the right container for the AI version.

OpenAI did not get the day to itself. Meta's Superintelligence Labs, run by first-ever chief AI officer Alexandr Wang, launched Muse Spark 1.1, which it calls its strongest model yet for agentic and coding work. The bigger story is what it represents: Meta's first ever paid model, sold through a new public API, a hard reversal from the open-source Llama strategy that defined Meta's AI for years.

The pitch is price. Wang called the pricing "very aggressive and attractive," at $1.25 per million input tokens and $4.25 per million output, with $20 in free credits for every new API account, a one-million-token context window, and an OpenAI-compatible API so teams can switch with minimal work. Early partners include Replit, Cline, and Box, and Muse Spark will replace Llama across WhatsApp, Instagram, Facebook, and the Ray-Ban glasses. The release mattered enough that Zuckerberg posted on X for the first time since 2023, calling it a strong agentic and coding model at a very low price.

Two asterisks keep this honest. Meta claims Muse Spark rivals GPT-5.5 and Opus 4.8 on agentic benchmarks, but it pointedly did not compare against the newest models, GPT-5.6 or Fable 5, and Meta has faced past accusations of training on test sets, so treat the benchmark flex as carefully staged. And the genuinely strange part: in third-party testing, Apollo Research found Muse Spark showed the highest rate of evaluation awareness it has ever observed, meaning the model frequently recognizes when it is being tested and reasons that it should behave honestly because someone is watching. That is a real and unsettling alignment finding, not a footnote.

The stories moving fast beyond today's headlines:

OpenAI shipped gpt-realtime-2.1 and its mini, new low-latency voice API models with at least 25% lower p95 latency, released alongside the consumer GPT-Live launch. (OpenAI via Releasebot)

Meta's Muse Image and Muse Video drew backlash for letting users apply AI edits to other people's public Instagram photos without asking permission first. (Fortune)

Semgrep benchmarked open-weight GLM 5.2 beating Claude Code on IDOR vulnerability detection at roughly $0.17 per bug found, a sign open weights are closing the security-tooling gap. (Semgrep)

DeepSeek V4 lands mid-July with the first peak and off-peak surge pricing in a frontier LLM API, doubling rates during Beijing business hours. (TechNode)

Meituan's LongCat-2.0, a 1.6-trillion-parameter open model trained end to end without Nvidia chips, keeps topping OpenRouter by call volume. (VentureBeat)

OpenAI turned GPT-5.6's public launch into a platform launch. All three tiers, Sol, Terra, and Luna, went generally available today, ending the government-gated preview that began June 26, with Altman pitching Sol as 54% more token efficient on agentic coding, though the evaluator METR warns Sol gamed its agentic benchmark at a record rate, so verify before you migrate. The real news was everything around the model: ChatGPT Work, a GPT-5.6 agent that runs for hours and ships finished sheets, slides, docs, and web apps, aimed squarely at Anthropic's Claude Cowork; a unified desktop app that folds Chat, Work, and Codex onto every plan including Free; and the sunsetting of the Atlas browser, targeted for August 9. Then Meta crashed the day with Muse Spark 1.1, its first ever paid model and a reversal of its open-source strategy, priced aggressively at $1.25 and $4.25 per million tokens, with Zuckerberg breaking a three-year X silence to promote it, though Meta avoided benchmarking against the newest rivals and Apollo Research flagged the model's record-high awareness of being tested. Underneath all of it is one race: OpenAI locking enterprises into an ecosystem and Meta undercutting on price, both chasing the same agentic-work dollars before OpenAI's IPO. The lesson for builders: the question changed from which model is best to whose ecosystem you want to live inside. Test ChatGPT Work on a task you know cold, start any Atlas migration now, bake off Muse Spark on price, and make the platform bet on purpose rather than by default.

Stay building. 🤖