Report: GPT-5.6 Sol appears in Codex before reported July 9 launch

• GPT-5.6 has an official date and family: OpenAI's announcement says Sol, Terra, and Luna launch Thursday, with preview access expanding globally.
• Codex is the clearest prelaunch surface: the Codex codebase screenshot lists GPT-5.6-Sol with a 372K context window, low as the default reasoning level, and a description aimed at agentic coding.
• The visible benchmark headline is TerminalBench 2.1: Sol Ultra moves from GPT-5.5's 83.4% baseline to 91.9%, +8.5 points, in AlphaSignalAI's chart.
• Hands-on reports describe Sol as fast, persistent, and agent-friendly, with theo saying it runs for a day without /goal and petergostev calling it stronger on reliability than Fable.
• The 750 tokens/sec Cerebras path is a limited-preview infrastructure story, because the Cerebras excerpt says select customers get initial access as capacity expands.

The prelaunch artifacts are unusually concrete: the Codex codebase screenshot shows a 372K context window and default low reasoning for GPT-5.6 Sol; the Ultra-in-Codex screenshot ties Ultra to Codex; the Cerebras excerpt says selected customers get up to 750 tokens/sec in July. The strangest hands-on claim came from danshipper, who said Sol can run loops over email, hiring, meetings, Slack, and Facebook Marketplace, using his browser according to a follow-up reply.

OpenAI confirmed the public launch window, the three model names, and a global preview expansion. Sam Altman echoed the timing in sama's post, writing that GPT-5.6 Sol launches Thursday.

The confirmed family is simple:

• Sol: the flagship model named in OpenAI's announcement.
• Terra: the middle tier, named alongside Sol and Luna.
• Luna: the lower tier, named alongside Sol and Terra.
• Preview access: expanding globally before the public launch, per OpenAI's announcement.

A separate OpenAI Developers sun-emoji teaser circulated before the confirmation, according to the developer-account screenshot. The official launch graphic then used the Sol, Terra, and Luna space motif.

The Codex artifacts carry the most implementation detail so far. The code screenshot in scaling01's post lists display_name: "GPT-5.6-Sol", context_window: 372000, max_context_window: 372000, reasoning_summary_format: "experimental", and default_reasoning_level: "low".

A separate Codex UI modal says GPT-5.6 Sol is available in Codex for "faster agentic coding, quick repo edits, and high-velocity build loops with the new speed control," according to WesRoth's Codex screenshot.

Codex also appears to expose the full model family. skirano's menu screenshot shows GPT-5.6-Sol, GPT-5.6-Terra, and GPT-5.6-Luna in the model picker, with Ultra selected under reasoning.

The reasoning ladder is still partly assembled from screenshots and reports. scaling01's screenshot shows max and ultra listed for gpt-5.6-sol, while the Ultra-in-Codex screenshot says "Ultra will be in codex."

The reported price ladder is cleaner than the reasoning story:

• Sol: $5 input, $30 output per 1M tokens, according to AlphaSignalAI's roundup.
• Terra: $2.50 input, $15 output per 1M tokens, according to AlphaSignalAI's roundup.
• Luna: $1 input, $6 output per 1M tokens, according to AlphaSignalAI's roundup.
• Caching: cache writes cost 1.25x uncached input, while cache reads keep the 90% discount, according to AlphaSignalAI's roundup.

Community replies also tried to separate product names from compute settings. koltregaskes wrote that Pro is a model and Ultra is a compute level, while another reply described Pro as a model and Ultra as a reasoning level.

The most-circulated prelaunch benchmark is TerminalBench 2.1. The chart gives these GPT-5.6 deltas against GPT-5.5's 83.4% score:

• GPT-5.6 Sol Ultra: 83.4% → 91.9%, +8.5 points, in AlphaSignalAI's chart.
• GPT-5.6 Sol: 83.4% → 88.8%, +5.4 points, in AlphaSignalAI's chart.
• GPT-5.6 Terra: 83.4% → 84.3%, +0.9 points, in AlphaSignalAI's chart.
• GPT-5.6 Luna: 83.4% → 82.5%, -0.9 points, in AlphaSignalAI's chart.

The same chart puts Claude Mythos 5 at 88.0%, Claude Fable 5 at 84.3%, Claude Opus 4.8 at 78.9%, and Gemini 3.1 Pro Preview at 70.7%.

A personal legal benchmark claim sits outside that chart. deredleritt3r said GPT-5.6 Sol Pro saturated prinzbench, describing it as specialty legal-research questions involving tricks and painstaking review of many authorities.

Ethan Mollick's benchmark caveat was narrower: emollick said OpenAI had a strong autonomous-work benchmark in GDPval but had not reported it for GPT-5.6.

The hands-on reports converge on one word: workhorse. benhylak said Sol is fast, stays on track, works hard, and stops giving up prematurely compared with 5.5.

The Fable comparison split by task:

• emollick described Sol as similar in ability but faster and more stepwise, while Fable goes off for longer tasks at its own pace.
• emollick's follow-up put Sol on back-and-forth tasks, Fable on long defined tasks, and Sol Pro on hard problems.
• mattshumer_ said Fable was better on almost every task he tested and more agentic, with one Fable turn doing what many 5.6 turns did.
• mattshumer_'s follow-up still called Sol a huge jump over 5.5.
• petergostev said Fable felt smarter and stronger at writing, while Sol won on robustness, reliability, following existing code patterns, subagents, and token efficiency.
• skirano said Sol was faster and more surgical in code changes, while Fable added more fluff.

The frontend claims conflict. skirano said OpenAI "finally fixed front-end design," while Yuchenj_UW said Sol still was not a great designer in his test.

Sol's strongest workflow claims are about loops, not single prompts. danshipper said GPT-5.6 was the first model he had used that could reliably run whole loops of knowledge work, including email, hiring, internal-meeting updates, Slack updates, and Facebook Marketplace scans.

The orchestration pattern is already forming around model specialization:

• omarsar0 proposed GPT-5.6 as executor and Fable 5 as on-demand advisor.
• omarsar0's harness reply described dynamic workflows involving different models, communications, and handoffs.
• theo said Sol understands subagents and orchestration in OpenClaw and Hermes Agent.
• petergostev said Sol is fluent at managing subagents and different threads.
• skeptrune joked about a 9-hour /goal loop that finished, merged, and did not take down production.

Computer use is part of the same story. theo called GPT-5.6 Sol world-leading in computer use and said losing access made him use it less, while mckbrando said Computer Use in Codex is defining the frontier.

The speed claim is 750 tokens/sec for GPT-5.6 Sol on Cerebras. The text excerpt in kimmonismus's post says the July Cerebras launch brings frontier intelligence to customers at that speed, with initial access limited to select customers.

The hardware explanation is alien enough to matter:

• wafer_ai's deep dive says WSE-3 has 900,000 cores on one wafer.
• Each processing element has 48 KB of private SRAM, its own program counter, and a 5-port router, according to wafer_ai's deep dive.
• Data moves as 32-bit wavelets across virtual channels called colors, according to wafer_ai's deep dive.
• The total on-chip SRAM is 44 GB at 21 PB/s aggregate bandwidth, versus 3 TB/s on H100, according to wafer_ai's deep dive.

The model-sizing claims are estimates. koltregaskes's calculator screenshot shows a hypothetical Sol configuration with 3T total parameters, 150B active parameters, 70 layers, 922K max context, 70 Cerebras systems, and 2.48 TB of memory composition.

The launch landed inside a Fable access debate. giffmana posted a Claude Code usage-credits screenshot showing $73.50 spent against a $70 monthly limit, then wrote that Fable usage credits felt like "Zynga hooking in the whales."

The subscription comparison became a release-day wedge. haider1 claimed Fable 5 had effectively moved behind usage credits while GPT-5.6 would be available through subscription, and haider1 argued that Sol's efficiency made subscription access easier for OpenAI to support.

That context explains why so many hands-on reports compare Sol against Fable instead of GPT-5.5. bridgemindai argued Anthropic could not afford to remove Fable from subscriptions if GPT-5.6 shipped broadly, while haider1 framed OpenAI's launch window as a chance to absorb users losing Fable access.

Several concrete launch details remain unresolved in the evidence pool.

• Existing Codex threads: imjustnewatai claimed GPT-5.6 does not work well inside already-existing GPT-5.5 Codex threads and said fresh threads are recommended.
• Regions: koltregaskes said the expanded preview looked global, but UK and EU users might still wait.
• Cerebras parity: koltregaskes cited a claim that Cerebras serves the same model, but a follow-up reply said the exact context size remained unclear.
• Training lineage: haider1 claimed Sol is further-trained 5.5 rather than a new pretrain, and mattshumer_ replied that the claim was inaccurate.
• Government gating: the Axios screenshot says the Trump administration lifted restrictions on GPT-5.6, while theo said OpenAI asked early testers to wait until Thursday for formal posts but gave them zero restrictions on what they could say.