GPT-5.6 Sol Ultra ships general availability

• GPT-5.6 Sol Ultra became generally available, and danshipper reported an accidental 2B-token overnight run within a day of launch.
• Creator tests jumped straight into artifacts: cinematic websites from viktoroddy's tutorial, a 3D Torii travel site from petergyang's frontend test, and 3D worlds from minchoi's roundup.
• The cleanest early win was orchestration: stevibe's canvas tests found GPT-5.6 stronger on multi-step UI behavior, while GPT-5.5 still won two raw physics prompts.
• The desktop app rollout merged Codex into ChatGPT Work/Codex, with danshipper's day-zero review calling it useful and thekitze's install screenshot showing the rebrand landing roughly.
• Day-one surfaces included Figma Make, where figma said GPT-5.6 produced stronger outputs from existing designs and better token efficiency.

viktoroddy posted a one-click prompt bundle for cinematic AI websites. Higgsfield started running Sol against Fable through Seedance 2.0 in one side-by-side video, while underwoodxie96 turned a GPT Image 2.0 still of 1990 Shanghai into a 15-second Seedance documentary scene. The funniest practical caveat came from petergyang, who said Sol High Fast burned through usage much faster than GPT-5.5 High Fast.

The public naming surfaced as three model tiers plus effort modes. One launch summary from Everlier put the apparent API pricing at:

• Sol, flagship: $5 input and $30 output per 1M tokens.
• Terra, balanced: $2.50 input and $15 output per 1M tokens.
• Luna, fast and cheap: $1 input and $6 output per 1M tokens.
• Ultra mode: 4 parallel agents for complex tasks.

The labels were not self-explanatory on day one. bentossell noted that “Sol light” on mobile mapped to “Sol low” on desktop, and petergyang asked when to use Sol Medium, Sol High, Terra, and app-specific defaults.

The creator lane opened with web design, not benchmark charts. viktoroddy recorded a 21-minute tutorial on building “premium, cinematic websites” with GPT-5.6 Sol, then shared the prompt pack as a one-click bundle.

petergyang compared GPT-5.6 and Fable on a Japan travel site. GPT-5.6 built a site with a 3D Torii gate, and Yang’s prompt note was specific: ask for “a 3D WebGL element to the hero” when the goal is animated 3D graphics.

Figma shipped the model into Make the same day. zoink called the result “very good,” and figma said early tests showed stronger outputs from existing designs plus greater token efficiency.

The best prompt evidence looked more like production art direction than chatbot prompting.

AmirMushich used GPT-5.6 image prompting for a Switzerland 2026 World Cup campaign visual with a strict design system:

• 75 percent universal composition across teams.
• 25 percent localized national color, material, motif, and emblem choices.
• Exactly two text elements: team name and “2026.”
• No players, match action, sponsor logos, FIFA wordmark, watermarks, or invented text.

underwoodxie96 showed a two-stage image-to-video workflow:

1. GPT Image 2.0 generated a documentary still of dock workers near Shanghai’s Huangpu River in 1990.
2. Seedance 2.0 animated the reference image into 15 seconds of footage.
3. The video prompt locked faces, clothing, location, film stock, handheld cuts, natural sound, and historical exclusions.

[minchoi's roundup] gathered the launch-day “look what it built” thread into a creator menu:

• Manhattan recreated in voxels, from minchoi's first example.
• A Google Earth clone with 3D terrain, cities, weather, day/night, and global search, from minchoi's Earth clone post.
• A one-prompt, playable flight simulator in 10 minutes, from minchoi's flight sim post.
• A phone turned into a wireless mic as an Android app, from minchoi's mobile app example.
• A Blender cannon build, from minchoi's Blender post.
• A 3D dashboard comparison where GPT-5.6 was “half the cost” and close to Fable, from minchoi's dashboard comparison.

Games were mixed but useful. petergyang got both GPT-5.6 and Fable to one-shot a 3D Star Fox-style level with enemies, power-ups, and a boss battle in about 10 minutes, then gave Fable the edge because it remembered barrel rolls.

Two smaller 3D details mattered for makers. kiaran_ritchie said Sol one-shot a GPU painting system that was “ultrafast” and robust across UV seams, while om_patel5 highlighted an open-source tool that turns a photo into an editable Three.js model written as code.

Higgsfield turned the Sol versus Fable debate into a video prompt shootout, with both outputs rendered through Seedance 2.0.

The comparison set covered:

• Side-by-side generated video, from higgsfield_ai's first comparison.
• Action scenes, from higgsfield_ai's action-scene test.
• Cartoon animation, from higgsfield_ai's cartoon round.
• Game POV output, from higgsfield_ai's game POV test.
• Paper-cut animation, from higgsfield_ai's paper-cut round.

The pattern was useful because it held the video renderer constant. The variable was the model directing the same brief.

stevibe ran four canvas animation prompts across GPT-5.6 Sol Ultra, Terra Ultra, Luna Max, and GPT-5.5 xHigh:

• Mac “hello”: Sol Ultra and GPT-5.5 tied for the most authentic cursive.
• Fireball into water: GPT-5.5 did better on steam burst and bubble plume.
• Burning wet paper: GPT-5.5 won again because flame stopped at the water edge and the wet half charred instead of burning.
• ChatGPT app UI: GPT-5.6 pulled ahead because all three variants followed the multi-step sequence and streamed the reply word by word.

The caveat showed up again in LLMJunky's Rocket League test, where GPT-5.6 Sol xHigh produced a “pretty good” demo, but Fable and GPT-5.5 Pro were described as more impressive on the same style of prompt.

The Every review framed GPT-5.6 as a daily driver rather than the biggest model in the room. danshipper said Sol scored 56/100 on Every’s Senior Engineer benchmark versus 91 for Fable, but called it the best writer among frontier models and the first model he trusted to run full loops of knowledge work.

Those loops were concrete. danshipper listed email, hiring, internal meeting and Slack decisions, and Facebook Marketplace scanning as recurring tasks Sol could run. gregisenberg's masterclass post expanded the workflow into inbox cards, Slack and meeting feeds, an agent email address, repeatable skills learned from one demonstration, and long goals running for up to 20 hours.

Builders reported similar agent stacks around Codex. MengTo said GPT-5.6 pushed him “all in on Codex,” citing agent skills, goals, spawned threads, browser use, computer use, and a mobile connector. alliekmiller called the model an “execution beast” and said it made bad execution, slow bug fixes, and unhelpful customer support harder to tolerate.

The app story was less clean than the model story.

The complaints clustered around four details:

• Codex became the ChatGPT app, while a checkbox let users keep the Codex icon, according to LLMJunky's icon note.
• Work and Codex appeared as modes, but thekitze said switching between them did not change the sidebar.
• danshipper pointed at the wording problem: Work was “for getting work done,” while Codex was “for developers.”
• petergyang's feedback post said the tabs, model tiers, effort settings, tasks, and chat history created confusion for regular users.

Sites also looked unfinished for some users. petergyang's Sites screenshot showed a publish attempt ending in a 404 and a message that Sites was not connected to the account.

GPT-5.6 was not only inside ChatGPT.

Day-one and near-day-one surfaces in the evidence included:

• Figma Make, where figma said GPT-5.6 was available starting launch day.
• Convos, where ShaneMac showed GPT-5.6 Terra as the default model for an agent draft.
• Hermes Agent, where ozansihay showed GPT-5.6 Terra and Luna alongside Grok 4.5.
• ChatGPT mobile, where bentossell showed a new “Remote” entry in the app menu.
• OpenClaw, where steipete said beta or dev channel code from after the OpenAI announcement was needed because of last-minute changes.

Speed claims also started appearing outside OpenAI. steipete said GPT-5.6 on Cerebras was “10x speed.”

The launch-day constraint was not access, it was consumption.

danshipper said he accidentally spent 2B tokens overnight after Sol Ultra became generally available. petergyang said he never thought about usage on GPT-5.5 High Fast, but Sol High Fast burned through usage quickly enough that he planned to test Sol Medium without Fast.

The settings question stayed open in public. petergyang asked whether Sol Medium or Sol High should be the default, when Terra beats Sol Medium, and what the best default is for tools like OpenClaw or Hermes. In the same thread, petergyang said Extra High “thinks too much.”