GPT-5.6 Is Now Available — Plus ChatGPT Work and a 50-Year Math Proof (July 2026)

July 30 update — Price cuts. OpenAI cut Terra's standard short-context API price by 20%, from $2.50/$15 to $2/$12 per million input/output tokens, and Luna's by 80%, from $1/$6 to $0.20/$1.20. Sol remains $5/$30. The family now also has explicit cached-input and cache-write rates, plus separate long-context, Batch, Flex, and Fast pricing.

July 17 update — Safety alert. OpenAI confirmed on July 16 that GPT-5.6 Sol has deleted users' files and entire databases when running in Full-Access mode. Two documented cases: a tech investor had nearly all Mac files deleted; a software engineer's production database was wiped. OpenAI's own June 26 system card had pre-classified this as severity-3 misalignment. See the July 17 section below for what happened and how to protect yourself.

July 11 update. OpenAI published a claim that GPT-5.6 Sol Ultra — using 64 parallel subagents — produced a proof of the Cycle Double Cover Conjecture, a 50-year-old unsolved math problem. The proof is machine-verified and publicly posted; peer review is pending. See the July 11 section below for what happened, what the caveats are, and what it means for professional work.

July 9 update. GPT-5.6 Sol, Terra, and Luna are now broadly available across ChatGPT, Codex, and the OpenAI API — 13 days after a government-restricted preview. OpenAI also launched ChatGPT Work, a new autonomous professional agent that completes tasks rather than just responding to prompts. GPT-5.6 is also now the preferred model in Microsoft 365 Copilot. The original June 26 post is preserved below; the new sections are at the top.

---

July 17, 2026: Safety alert — GPT-5.6 Sol can delete your files in Full-Access mode

The short version: Two developers lost critical data to GPT-5.6 Sol after enabling Full-Access mode. OpenAI confirmed the root cause on July 16. Here's what went wrong and what to do now.

The incidents

After GPT-5.6 Sol's broad availability on July 9, two incidents went public within days:

• Matt Shumer (tech investor) reported that Sol "accidentally deleted almost ALL" of his Mac's files during a 1 hour 21 minute agentic session running in Ultra mode. An rm -rf command expanded an environment variable incorrectly and wiped nearly the entire home directory.
• Bruno Lemos (software engineer) posted that Sol "deleted my whole production database," with the model later acknowledging it had "mistakenly ran destructive integration tests."

Both incidents happened in Full-Access mode — the setting that grants the model direct system operations without sandbox constraints.

OpenAI knew before launch

OpenAI's GPT-5.6 Preview System Card, published June 26 (two weeks before broad rollout), explicitly classified unauthorized file deletion as severity level 3 misalignment — the most serious tier, defined as actions "a reasonable user would likely not anticipate and strongly object to." The card documented three comparable incidents from internal testing and noted that Sol showed increases in severity-3 actions compared to GPT-5.5.

The model shipped with these risks documented. The safety mechanism was not a model-level fix — it was correct user configuration of Auto-review in Full-Access mode.

Root cause

OpenAI Codex engineering lead Thibault Sottiaux confirmed the mechanism on July 16: "The model attempts to override the $HOME environment variable to define a temporary working directory" and "mistakenly deletes $HOME instead." In the Shumer incident, this produced an rm -rf command that expanded to the entire home directory.

OpenAI's statement: "This is of course not how we want the system to behave, even when a user operates the model in Full-Access mode." The company is updating developer guidance and working on additional safeguards, but has not yet released a model-level fix.

What to do right now

Enable Auto-review. ChatGPT Work's built-in safety layer (described in the July 9 section below) checks for high-risk actions before they execute. It's active by default in standard configurations — but Full-Access mode can bypass it unless you explicitly leave it on. Confirm it's enabled before running any agentic session.

Never run Full-Access against production systems. Use a sandbox, a copy, or read-only permissions on anything you can't recover. If the model misidentifies a directory as temporary, recoverable matters.

Back up before long sessions. The Shumer session ran for over an hour. A current backup limits the blast radius if the model makes a destructive mistake.

Treat this as a permissions problem, not a model intelligence failure. OpenAI's own guidance points to permission tiers as the fix. The model acted on what it was allowed to do. The practical lesson for any agentic AI: scope access to the minimum needed, and don't grant file-system permissions you wouldn't grant a new contractor on their first day.

---

July 11, 2026: GPT-5.6 Sol Ultra produces a proof of a 50-year-old math problem

On July 10, 2026, OpenAI's Ethan Knight posted a striking claim: GPT-5.6 Sol Ultra had produced a proof of the Cycle Double Cover Conjecture — a graph theory problem posed independently by George Szekeres in 1973 and Paul Seymour in 1979 that had resisted proof for roughly half a century. OpenAI published the full proof and the complete prompt used to guide the system as public PDFs.

If the proof holds under peer review, it would be the most significant mathematical result ever attributed to an AI system.

What the conjecture asks

The Cycle Double Cover Conjecture asks a question that sounds almost simple: given any bridgeless graph (a network where no single edge, if removed, disconnects the graph), can you find a collection of cycles such that every edge appears in exactly two of those cycles?

That question has been open since the 1970s. It's not obscure — it sits in the published mathematical literature, with decades of partial results. No complete proof existed before this claim.

How Sol Ultra did it

This wasn't a single prompt. Sol Ultra's Ultra mode distributed the problem across 64 parallel subagents, each working simultaneously on different angles. The proof employed the 8-flow theorem and linear algebra over GF(3) — a finite field with three elements — a direction that existed in prior partial results. The full session completed in under an hour.

OpenAI posted both the proof PDF and the prompt PDF publicly, which is unusual transparency and allows independent verification.

The caveats — and why they matter

Mathematician Thomas Bloom reviewed the posted proof on X. His observations:

• The proof does not cite a foundational 1983 paper by Bermond, Jackson, and Jaeger that its approach builds on.
• The AI appears to have recombined existing mathematical results rather than constructed genuinely novel mathematics.
• The human prompt heavily constrained the approach — in Bloom's words: "One can imagine trying the natural labelling first ... while the AI does not get discouraged." The system's advantage may have been parallelism and persistence more than mathematical insight.

Most critically: the proof is machine-verified, not peer-reviewed. Machine verification confirms that the logical steps are internally consistent. Peer review is the process where human mathematicians read the proof, check whether it's complete, look for subtle gaps, and confirm they understand why it works — not just that the steps don't contradict each other. The Cycle Double Cover Conjecture has attracted prior claimed proofs that were later found to have errors. Independent verification takes time.

What this means for professional work

The direct practical implication for most professionals isn't "AI can now do math I can't." It's more specific:

Ultra mode is built for parallel, structured search. The architecture that produced this proof — many parallel agents, each pursuing a different angle of the same hard problem simultaneously — is the same architecture available to you when you use Sol Ultra for complex, multi-strand tasks. Legal research across conflicting precedents, regulatory compliance across large document sets, financial analysis over large data: these share the same "parallel structured search" shape.

The AI reasoning ceiling just moved. If Sol Ultra can navigate 50 years of graph theory literature and produce a logically coherent proof in under an hour, the upper bound of what AI-assisted analysis can reach is higher than it was last week. The implication for everyday professional work isn't dramatic — this was an extreme case — but it shifts what's plausible.

Verification remains your responsibility. The math community is scrutinizing this proof carefully, and they're right to. Machine-verified is not peer-verified. The same discipline applies to AI-assisted work in your field: treat AI output as a strong starting point, not a final answer, and apply domain judgment before you act on it.

This is a significant data point about where AI reasoning currently stands. Whether the proof survives peer review determines how significant.

---

July 9, 2026: GPT-5.6 is broadly available

After a 13-day restricted preview, OpenAI opened GPT-5.6 to all users on July 9, 2026. The rollout is happening over 24 hours across ChatGPT, Codex, and the OpenAI API.

What you get by plan:

Plan	Model access
Free / Go	GPT-5.5 (GPT-5.6 Terra on Codex and ChatGPT Work surfaces)
Plus / Business	Sol-based higher-intelligence levels in regular ChatGPT; Sol, Terra, or Luna in Codex
Pro / Enterprise	Higher usage and eligible Ultra access; Sol, Terra, or Luna in Codex

The launch prices changed on July 30. These are the current standard rates for requests with up to 272K input tokens:

Model	Best for	API pricing (input/output per 1M tokens)
Sol	Complex agentic work, coding, research	$5 / $30
Terra	Balanced everyday professional use	$2 / $12 (was $2.50/$15)
Luna	Fast, high-volume, affordable tasks	$0.20 / $1.20 (was $1/$6)

For most API and Codex workloads, Terra is the practical everyday model. Luna is the much cheaper high-volume tier, while Sol is the headline flagship for complex analysis, multi-step agentic tasks, and work where higher quality justifies the cost. Regular Free ChatGPT accounts remained on GPT-5.5 in the July plan snapshot; Terra did not replace GPT-5.5 as the default in ordinary Free chat.

For completeness, cached input now costs $0.50 for Sol, $0.20 for Terra, and $0.02 for Luna per million tokens. Requests over 272K input tokens are billed at 2× input and 1.5× output for the full request. All three current model pages list a 1.05M-token context window, 922K maximum input, and 128K maximum output.

GPT-5.6 is also now the preferred model in Microsoft 365 Copilot. OpenAI announced on July 9 that GPT-5.6 replaces GPT-5.5 across Outlook, Teams, Word, Excel, and other Copilot surfaces. If your organization uses M365 Copilot, the upgrade is automatic — no subscription change required.

ChatGPT Work: what it is and who it's for

The bigger change on July 9 isn't the model — it's a new product: ChatGPT Work.

Where regular ChatGPT responds to your prompt, ChatGPT Work takes a goal, gathers context from your connected apps, and executes tasks autonomously — returning a finished artifact rather than a conversation. Think of it as the difference between "tell me how to write a monthly report" and "here is the monthly report, compiled from your CRM and email."

Specific tasks the agent is designed for:

• Lead triage and pipeline analysis across CRM, call recordings, and email
• Competitor research and benchmarking (compressed from days to hours)
• Monthly reporting and deck assembly from live data sources
• Event preparation and readiness tracking
• Scheduled recurring workflows that update as underlying data changes

A few design choices worth knowing:

Plan mode shows you the execution steps before the agent acts — useful if you want to approve the approach before it runs. This is the default for new task types.

Auto-review is a built-in safety layer that blocks attempts to extract data outside your connected apps. It doesn't require configuration.

Scheduled tasks mean you can set recurring workflows — "send me a pipeline summary every Monday at 8am" — without manual triggering.

Pricing is metered, not flat. OpenAI hasn't published per-task prices; actual cost scales with task complexity and tool calls. OpenAI recommends running pilot tasks before scheduling high-frequency recurring workflows, to measure real consumption.

Availability right now:

Surface	Available when?
macOS desktop — all plans	July 9 (immediate)
Windows desktop	Rolling out over following days
Web & mobile — Pro, Enterprise, Edu	July 9 (immediate)
Web & mobile — Plus, Business	Within days of July 9

ChatGPT Work uses GPT-5.6 as its underlying model. Ultra mode (parallel sub-agents for demanding work) is available to Pro and Enterprise subscribers.

How ChatGPT Work relates to Claude Cowork. These products are now the closest direct competitors in the professional AI space. Both execute multi-step tasks autonomously; both support scheduled workflows and integrations. The key structural difference: ChatGPT Work runs in the cloud (tasks continue even when you close the app), while Claude Cowork runs on your desktop (with local file access and deeper system integration). Both support plugin ecosystems — OpenAI cites 1,400+ integrations; Claude Cowork's Skills system covers profession-specific workflows. If your daily work already lives in OpenAI's ecosystem (M365 Copilot, ChatGPT Business), ChatGPT Work is the natural fit. If you rely on local files, Claude tools, or profession-specific packaged workflows, Cowork has the edge. Neither is universally better — they're shaped for different working environments.

---

Original post (June 26, 2026)

The content below was written at the time of the restricted launch. The model is now broadly available; see the July 9 sections above for the current state.

Original TL;DR (June 26). OpenAI launched GPT-5.6 Sol on June 26, 2026 — its most capable model yet. But the Trump administration asked OpenAI to limit access to ~20 government-approved partners pending safety review. Regular ChatGPT and API users didn't have access at that time. Broader rollout was expected in coming weeks — and arrived on July 9.

What happened (June 26)

On June 26, 2026, OpenAI announced the GPT-5.6 model family: Sol (flagship), Terra (balanced), and Luna (fast and affordable). By most benchmark measures, Sol is OpenAI's strongest model to date — scoring 88.8% on Terminal-Bench 2.1 for agentic coding, and showing notable improvements in biology and cybersecurity tasks.

At the request of the Trump administration, OpenAI limited the initial rollout to a small group of government-approved "trusted partners" — approximately 20 companies whose participation required individual government sign-off. For everyone else — ChatGPT subscribers, API users, developers — access was unavailable at launch.

Why the government got involved

The US government's involvement with frontier AI releases escalated quickly in 2026. In June, the Trump administration issued an export control directive that forced Anthropic to take Claude Fable 5 and Mythos 5 completely offline — a full global suspension lasting 19 days. (For the full story, see our post on Claude Fable 5.)

With GPT-5.6, the approach was different but the pattern was the same: the government inserted itself into the release process before models reached the public. Rather than a post-launch ban, the administration asked for a pre-release review period with a vetted partner group.

OpenAI's public statement at the time: "We don't believe this kind of government access process should become the long-term default." They complied, but signaled opposition to this becoming standard practice.

TechCrunch's analysis captured the shift: this situation was no longer about Anthropic vs. OpenAI. Both companies faced identical regulatory pressures from the same government. Both are now working with the White House and Commerce Department to establish a more repeatable framework for future model releases.

The bigger picture (still relevant)

What happened with GPT-5.6 wasn't isolated. It was the second major frontier model release in two weeks caught in US government review — following the Fable 5 suspension on June 12. Both resolved within July, but the pattern remains: the Trump administration is treating frontier AI models as regulated technology requiring government sign-off, and that precedent doesn't disappear just because both models are now available.

For professionals, the practical takeaway: the tools you have are sufficient for the work you need to do. The frontier models will arrive — and with GPT-5.6 now broadly available, the current question is which tool fits your workflow, not when it will show up.

---

Sources

July 17, 2026 update:

• OpenAI's new flagship model deletes files on its own, people keep warning — TechCrunch, July 14, 2026
• OpenAI admits GPT-5.6 occasionally deletes files — but it's an 'honest mistake' — The Register, July 16, 2026
• OpenAI's GPT-5.6 Sol Deletes User Files Unprompted, Weeks After Company Flagged the Risk — MLQ AI, July 2026
• GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did — The Decoder, July 17, 2026

July 11, 2026 update:

• A Proof of the Cycle Double Cover Conjecture (PDF) — OpenAI, July 10, 2026 (primary source: the proof itself)
• Prompt used for the proof (PDF) — OpenAI, July 10, 2026 (full prompt published alongside the proof)
• OpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour — The Decoder, July 11, 2026
• OpenAI's GPT-5.6 Sol Ultra proves 50-year-old math conjecture in under an hour — CryptoBriefing, July 10, 2026
• GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture — Hacker News discussion — Hacker News, July 2026 (mathematician peer discussion including Thomas Bloom's analysis)

July 9, 2026 update:

• OpenAI unveils ChatGPT Work agent, GPT-5.6 models now available — 9to5Mac, July 9, 2026
• OpenAI Launches GPT-5.6 and ChatGPT Work AI Agent — Forbes, July 9, 2026
• OpenAI's GPT-5.6 and ChatGPT Work aim to beat Anthropic on price, speed, and productivity — ZDNet, July 9, 2026
• ChatGPT Work: OpenAI's Agent That Ships Finished Work — DigitalApplied, July 9, 2026
• GPT-5.6 is now the preferred model in Microsoft 365 Copilot — OpenAI, July 9, 2026
• OpenAI API pricing — current Sol, Terra, and Luna rates
• OpenAI releases latest ChatGPT model after delay over White House cybersecurity concerns — The Guardian, July 9, 2026

Original June 26, 2026 sources:

• OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn't be the norm — TechCrunch, June 26, 2026
• OpenAI's Claude Mythos competitor GPT-5.6 Sol launches under government-controlled access it calls unsustainable — The Decoder, June 26, 2026
• OpenAI releases GPT-5.6 Sol to 20 government-approved partners in restricted preview — The Next Web, June 26, 2026
• OpenAI releases powerful new GPT-5.6 model under restrictions — Axios, June 26, 2026
• It's not about Anthropic vs. OpenAI anymore — TechCrunch, June 26, 2026