OpenAI releases GPT-5.6 Sol, its most powerful AI model yet

OpenAI rolls out GPT-5.6 family as frontier-model race tightens with Anthropic

OpenAI publicly released its new flagship artificial-intelligence model, GPT-5.6 Sol, on Thursday, alongside two lower-cost variants and a workplace agent, in the company’s most direct challenge yet to rival Anthropic. The New York Times reported that the rollout “escalates the battle for A.I. dominance” with Anthropic as the technology grows increasingly sophisticated.

The GPT-5.6 family ships in three tiers: Sol as the workhorse flagship, Terra as a mid-range option, and Luna as a budget-friendly entry, TechCrunch reported, citing OpenAI. The models are available across ChatGPT, Codex and the OpenAI API at $5/$30 per million input/output tokens for Sol, $2.50/$15 for Terra and $1/$6 for Luna, according to the same report.

Alongside the models, OpenAI unveiled ChatGPT Work, an AI agent that can operate spreadsheets, calendars and email on a user’s behalf, running on desktop, web and mobile. The New York Times described it as similar in concept to Anthropic’s Claude Cowork, while TechCrunch said it is “designed as a workplace companion for enterprise teams” for drafting documents, spreadsheets and presentations.

A model cleared by Washington, after weeks of friction

GPT-5.6 Sol’s public launch came after a stretch of tension with the U.S. government. The New York Times reported that the Trump administration initially treated OpenAI’s and Anthropic’s frontier models as national-security concerns, signing an executive order last month asking tech companies to voluntarily give the government oversight of new AI models before release. “Last week, OpenAI said it would share its new technology with only a small group of companies approved by the Trump administration,” the Times wrote, “But after lengthy discussions, the administration allowed both OpenAI and Anthropic to release their new technologies publicly.”

The Hindu added that Anthropic had separately been treated as a supply-chain risk, forcing it to suspend access to Fable 5 and Mythos 5 for 19 days between June 12 and July 1, 2026, before the Commerce Department lifted the relevant controls. Independent security experts quoted by the Times said the cybersecurity dangers were overstated.

OpenAI is taking a more permissive stance than Anthropic on cyber capabilities. The Times reported that GPT-5.6 Sol “includes fewer guardrails, so that individuals and businesses can use the technology to defend themselves,” while OpenAI separately called Sol “its strongest cybersecurity model yet, achieving frontier performance with significantly fewer tokens,” per TechCrunch. GPT-5.6 supports threat modeling, code review, patching and blue-teaming, TechCrunch said.

Benchmarks: where the two sides agree, and where they don’t

The benchmark fight is messier than either company claims. OpenAI told TechCrunch that Sol sets a new state of the art on the Artificial Analysis Coding Agent Index at 80, “2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less,” with Terra just above Fable 5 and Luna outperforming Anthropic’s Opus 4.8.

The Hindu, however, pointed to an independent counter-data point: Anthropic’s Fable 5 still leads on SWE-Bench Pro, a closely watched agentic-coding benchmark, at 80.3% versus the 64.6% OpenAI reports for Sol — roughly a 15-point gap. Vals AI chief executive Rayan Krishnan told the Times that on standard benchmark tests, OpenAI’s new flagship “roughly matches the performance of Anthropic’s leading model, Fable 5,” and that in real-world financial and legal tasks, “it is state of the art.”

The Hindu offered a structural explanation for the split: SWE-Bench Pro rewards careful, contained fixes to unfamiliar codebases, while Terminal-Bench-style and Coding-Agent-Index benchmarks reward broader workflow orchestration that suits Sol’s Codex-style harness. “A model can be a meticulous bug-fixer without being a great all-around operator, and vice versa,” the paper wrote — a useful framing for readers trying to make sense of competing SOTA claims.

On other benchmarks, OpenAI claims state-of-the-art marks of 92.2% on BrowseComp and 62.6% on OSWorld 2.0, and 53.6 on Agents’ Last Exam — “13.1 points ahead of Fable 5,” the company said.

Why it matters

The release is material for two reasons: enterprise buying decisions, and OpenAI’s financial trajectory. Barron’s framed the launch as potentially “help[ing] put the company back in the lead of the AI race before its expected initial public offering.” Pricing signals competitiveness: Krishnan told the Times that OpenAI’s top model is “more expensive than Fable in some cases,” but Sol’s per-token efficiency on coding tasks narrows that gap, and Luna at $1/$6 per million tokens undercuts Anthropic’s mid-tier.

The cybersecurity posture is also a competitive variable. Anthropic has layered restrictions onto Fable that block responses on cybersecurity, biology and other sensitive areas, routing those queries to Claude Opus 4.8. The Times cautioned that while those guardrails “may prevent hackers from using the technology for offensive purposes,” they “may also bar businesses and governments from using the system for defense,” an argument OpenAI is explicitly trying to capture with Sol.

Meta also released a new AI model on Thursday, the Times reported, “less powerful than GPT-5.6 Sol or Fable, and … considerably less expensive.”

Where the reporting diverges

The outlets converge on the basic facts — three model tiers, ChatGPT Work, benchmark claims — but diverge in tone. The New York Times emphasises the geopolitical backdrop and the cybersecurity trade-off between offense and defense. TechCrunch leans into OpenAI’s competitive positioning against Anthropic and provides the exact pricing and benchmark arithmetic. The Hindu drills into the mechanics of why both companies can claim wins, and is alone in quantifying the SWE-Bench Pro gap (80.3% vs. 64.6%) and the 19-day Anthropic outage. Barron’s, its body text behind a paywall, emphasises the IPO angle and a four-company stack-up that includes SpaceX and Google.

Independent verification is partial. TechCrunch explicitly cites the Artificial Analysis Coding Agent Index and OpenAI’s own in-house Codex toolkit; The Hindu notes that OpenAI evaluated Sol using its own Codex harness, which may inflate its lead on tooling-tuned benchmarks. Vals AI provides an outside read, but the Times’s Krishnan quote is one of the few independent assessments on record in these sources.

What to watch next

Three near-term signals will determine whether GPT-5.6 Sol’s launch is a turning point or a head-to-head draw. First, independent replication of the Artificial Analysis Coding Agent Index result, especially under standardised harnesses rather than OpenAI’s Codex tooling, would either validate or narrow Sol’s claimed 2.8-point edge.

Second, any further action from the Trump administration on frontier-model oversight will determine how durable Thursday’s clearance proves to be; The Hindu reported that the Commerce Department lifted controls on Anthropic only on July 1.

Third, enterprise procurement behaviour in the coming quarter — particularly among the kinds of large customers Anthropic has courted — will show whether the more permissive cyber posture of Sol or the narrower guardrails of Fable win in markets where compliance teams have a veto. The Times noted the OpenAI-Microsoft copyright lawsuit separately, a long-running issue that will continue to shadow any release from the company.