OpenAI says GPT-5.6 Sol launches publicly Thursday

• OpenAI put GPT-5.6 Sol, Terra, and Luna on a Thursday public-launch clock in its announcement, with preview access expanding globally now.
• Sol's early tester story is agent persistence: Theo said in his hands-on post that it can run for a day without /goal, understands subagents, and fixed his GPT-5.5 complaints.
• The visible benchmark win is TerminalBench 2.1: Sol Ultra scores 91.9% versus GPT-5.5 at 83.4%, +8.5 points, in the TerminalBench chart.
• Cerebras is the hardware swing: haider1's screenshot of OpenAI text says Sol will run at up to 750 tokens/sec, while the WSE breakdown explains the non-GPU programming model underneath.
• Fable remains the live comparison: Matt Shumer said in his early-access note that Fable was better on almost every task, while Theo called Fable smarter but Sol his default in his hands-on post.

OpenAI's announcement is a single space image with Sol, Terra, Luna, and "Coming Thursday" in the launch post. Codex screenshots already show GPT-5.6-Sol/Terra/Luna, Ultra reasoning, and a speed selector in skirano's UI screenshot and Wes Roth's Codex screenshot. The strangest detail is Cerebras's Wafer-Scale Engine: 900,000 processing elements, 48 KB SRAM per PE, 32-bit wavelets, and virtual channels called colors, according to the WSE breakdown.

• GPT-5.6 Sol, Terra, and Luna go public Thursday, according to OpenAI's announcement.
• Preview access is expanding globally now, according to OpenAI's announcement.
• Sol is the top tier, Terra is the balanced tier, and Luna is the lower-cost volume tier, according to AlphaSignalAI's roundup.
• Pricing is Sol at $5 input and $30 output, Terra at $2.50 input and $15 output, and Luna at $1 input and $6 output per 1M tokens, according to AlphaSignalAI's roundup.
• Cache writes cost 1.25x uncached input and cache reads keep a 90% discount, according to AlphaSignalAI's roundup.
• Sol Ultra appears as a reasoning level in Codex in skirano's UI screenshot.
• The rollout followed a restriction reversal: ai_for_success's Axios screenshot says the Trump administration lifted restrictions on GPT-5.6, and Theo later described government-related setbacks in his early-access note.

The chart attached to AlphaSignalAI's roundup shows one OpenAI-branded benchmark, TerminalBench 2.1.

First-party

• TerminalBench 2.1: GPT-5.5 83.4% → GPT-5.6 Sol Ultra 91.9%, +8.5 points, per the TerminalBench chart.
• TerminalBench 2.1: GPT-5.5 83.4% → GPT-5.6 Sol 88.8%, +5.4 points, per the TerminalBench chart.
• TerminalBench 2.1: GPT-5.5 83.4% → GPT-5.6 Terra 84.3%, +0.9 points, per the TerminalBench chart.

Third-party evaluators

No public GPT-5.6 third-party delta appears in the evidence pool. The prelaunch comparison target was Fable 5, which Kimmonismus's Artificial Analysis screenshot showed leading six occupational indices: Finance, Legal, Healthcare, Strategy & Ops, Engineering, and Economics.

Customer-reported

• CUDA speedup run time: Claude Opus 64-hour run → Sol surpassed it after 30 hours, 34+ hours sooner to surpass, according to Kimmonismus's screenshot of Bryce's post.

• TerminalBench 2.1: GPT-5.5 83.4% → GPT-5.6 Luna 82.5%, -0.9 points, in the OpenAI chart.
• Fable comparison: Matt Shumer said in his early-access note that Fable was "quite a bit better" for almost every task and more agentic.
• Shumer's follow-up still called Sol a huge jump over GPT-5.5 in his follow-up.
• Theo's caveat was similar but milder: his hands-on post called Sol less smart than Fable, more capable than GPT-5.5, and rough around fewer edges.

Cerebras serving is the launch's nerd Christmas detail.

• Wafer-Scale Engine: 900,000 processing elements in a 2D mesh, each with 48 KB private SRAM, its own program counter, and a 5-port router, according to the WSE breakdown.
• Programming model: dataflow, with 32-bit messages called wavelets moving along virtual channels called colors, according to the WSE breakdown.
• Memory model: 44 GB aggregate on-chip SRAM at 21 PB/s bandwidth, versus the H100 comparison of 80 GB HBM at 3 TB/s in the WSE breakdown.
• Sol estimate: 2T to 4T total parameters, about 150B active, 70 to 90 layers, and 70 to 100 Cerebras wafers, according to koltregaskes's Bleys summary.
• Open question: koltregaskes's same-model screenshot says Ahmed called the Cerebras version "the SAME model," while the same screenshot leaves context length unanswered.

No public system card in the evidence confirms whether Sol is a new base model, a finetune, or a hardware-shaped variant. OpenAI's announcement only names the family, the launch day, and the preview expansion.

• Codex UI: Wes Roth's screenshot shows a modal for "Introducing GPT-5.6 Sol" with speed control and rate-limit reset messaging.
• Codex feature flag: chetaslua's video shows GPT-5.6 UI elements becoming visible, while the post says direct model use was still limited to trusted users.
• Codex Ultra: Kimmonismus's Tibo screenshot says "Ultra will be in codex."
• Codex Mac app: Ben Hylak said in his Codex Mac note that GPT-5.6 felt like a natural fit there.
• API and Codex preview: AlphaSignalAI's roundup says selected trusted partners get preview access in API and Codex.
• Cerebras: Kimmonismus's OpenAI-text screenshot says Sol on Cerebras starts with select customers in July while capacity expands.

Claim: Sol beats Fable. Cited by: rohanpaul_ai's post called it the model that beats Claude Mythos, and the TerminalBench chart put Sol Ultra ahead of Claude Mythos 5 and Fable 5. Counter: Matt Shumer said in his early-access note that Fable was better on almost every task, while emollick said in his task split that he used Sol for back-and-forth work, Fable for very long tasks, and Sol Pro for hard problems. Evidence so far: TerminalBench favors Sol Ultra; hands-on testers split by workflow.

Claim: Cerebras 750 tokens/sec is the same Sol model. Cited by: koltregaskes's same-model screenshot quotes Ahmed saying "It's the SAME model." Counter: context length parity remains open in koltregaskes's follow-up. Evidence so far: the OpenAI-text screenshot says "up to 750 tokens per second" and initial access for select customers.

Claim: Sol's Thursday timing pressures Fable subscriptions. Cited by: haider1's timing post argued that releasing when Fable access tightens would pull demand toward OpenAI, and giffmana's usage-credit screenshot showed a Claude usage-credit limit already exhausted. Counter: rohanpaul_ai's newsletter screenshot says Anthropic extended Fable 5 paid-plan access through July 12, while aibuilderclub_'s note says the extension did not reset usage limits. Evidence so far: the access fight is quotas and credits as much as benchmark scores.

• Theo described Sol as determined, subagent-aware, good at orchestration, useful in OpenClaw and Hermes Agent, and strong on iOS development in his hands-on post.
• Theo later called Sol "world leading in computer use" in his computer-use note.
• Ben Hylak said in his early-access post that Sol is fast, stays on track, works hard, and gives up less prematurely than GPT-5.5.
• Matthew Berman said in his early-access post that GPT-5.6 feels like more than a dot-iteration improvement.
• emollick described Sol as faster and more stepwise than Fable in his early-tester note.
• skirano said in his hands-on post that GPT-5.6 fixed front-end design and that he had not needed to check generated code in two months.
• skirano's coding preference was surgical edits: his Fable comparison said Sol made tighter code changes while Fable added more fluff.
• Early-access comms stayed unusually open, according to Theo: his note on OpenAI says testers were asked to wait until Thursday for formal blog posts, videos, and podcasts, but had zero restrictions on what they could say.