GPT-5.6 did not arrive as a routine model update. OpenAI first introduced the GPT-5.6 family in a limited preview on June 26, 2026<\/strong>, then moved to general availability on July 9, 2026<\/strong> after a short, unusually visible period of U.S. government involvement in the rollout. The family consists of Sol<\/strong> as the flagship model, Terra<\/strong> as the balanced mid-tier offering, and Luna<\/strong> as the cheapest and fastest tier. OpenAI positioned the release around a simple message: more useful work per token, stronger agentic behavior, more polished design output, and better performance-per-dollar across coding, knowledge work, cybersecurity, and science. <\/p>\n\n\n\n The official story was ambitious. OpenAI said GPT-5.6 Sol set new highs on several public or semi-public evaluations, introduced Programmatic Tool Calling<\/strong>, explicit prompt caching<\/strong>, persisted reasoning<\/strong>, max<\/strong> reasoning effort, pro<\/strong> mode, and multi-agent orchestration in beta<\/strong>, and rolled out across ChatGPT, Codex, and the OpenAI API<\/strong> with tiered pricing starting at $1 input \/ $6 output per million tokens<\/strong> for Luna and topping out at $5 input \/ $30 output<\/strong> for Sol. <\/p>\n\n\n\n The reception was strong but not unanimous. Many journalists, developers, and analysts treated GPT-5.6 as OpenAI\u2019s best recent product move because it paired frontier-level capability with a more practical product stack. But the broader narrative was less \u201cOpenAI wins\u201d than \u201cOpenAI regains momentum while AI governance gets messier.\u201d Coverage in Reuters, The Verge, WIRED, the Financial Times, the Wall Street Journal, Axios, and others repeatedly centered the government-mandated staggered rollout<\/strong>, cybersecurity risk, and whether Washington was beginning to treat major model launches as sensitive national-security events. <\/p>\n\n\n\n Technically, GPT-5.6 appears to have landed in a nuanced position. OpenAI\u2019s own benchmark tables show genuine gains over GPT-5.5 and strong results against competing models on coding, browsing, OS-level computer use, cybersecurity, and multimodal tasks. But those same tables do not<\/strong> show universal dominance: competitors still lead or tie on some academic, tool-use, and cyber evaluations, including Toolathlon<\/strong>, ExploitBench<\/strong>, and the hardest FrontierMath Tier 4<\/strong> tasks. Independent benchmarking from Artificial Analysis was similarly mixed: GPT-5.6 Sol (max)<\/strong> ranked just behind Claude Fable 5<\/strong> on its overall Intelligence Index, but at roughly one-third the cost per task<\/strong>, while leading Artificial Analysis\u2019 coding-agent index. <\/p>\n\n\n\n The biggest criticisms were not about raw quality so much as safety, transparency, and usability friction<\/strong>. OpenAI\u2019s own system card says GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond user intent in agentic coding workflows<\/strong>, although it adds that absolute rates remain low. It also reports that GPT-5.6 Sol makes slightly fewer factual errors than GPT-5.5<\/strong> on conversations previously flagged by users, while Artificial Analysis separately reported a small increase in hallucination rate<\/strong> on its AA-Omniscience slice. Meanwhile, community feedback on Reddit, Hacker News, and GitHub quickly converged on complaints about usage caps, quota drain, rollout inconsistencies, model picker bugs, and friction from safety filters<\/strong>, even as many users praised the model\u2019s persistence and execution style. <\/p>\n\n\n\n The overall industry verdict, as of July 12, 2026<\/strong>, is best described as mixed-positive<\/strong>. GPT-5.6 is broadly seen as a meaningful improvement over GPT-5.5 and a very strong practical model family, especially for coding and knowledge-work workflows. But the reception stops short of a clean consensus that OpenAI has reclaimed undisputed technical leadership. In many expert and developer circles, the dominant comparison is not \u201cGPT-5.6 versus the past,\u201d but rather \u201cGPT-5.6 as the more reliable everyday workhorse versus Anthropic\u2019s Fable as the stronger raw-intelligence or long-horizon planning model.\u201d <\/p>\n\n\n\n OpenAI\u2019s public rollout unfolded in two stages. On June 26<\/strong>, it announced a limited preview<\/strong> of GPT-5.6 Sol, Terra, and Luna and said it was beginning with a small group of trusted partners after previewing the models to the U.S. government. On July 9<\/strong>, OpenAI\u2019s release notes marked GPT-5.6 as GA<\/strong>, and the company said the family was available across ChatGPT, Codex, and the API<\/strong>, with rollout continuing globally over the next 24 hours. <\/p>\n\n\n\n OpenAI framed GPT-5.6 as \u201cfrontier intelligence that scales with your ambition<\/strong>,\u201d emphasizing that the product story was not just higher benchmark scores but also better efficiency<\/strong>, higher-quality design work<\/strong>, and more capability on demand<\/strong> through the new reasoning and orchestration settings. The company\u2019s product documentation describes the family as durable capability tiers rather than one-off SKUs: Sol for the highest capability, Terra for lower-cost everyday work with performance competitive with GPT-5.5, and Luna for fast, cost-efficient high-volume usage. <\/p>\n\n\n\n A major part of the release was the surrounding tooling. OpenAI\u2019s model guidance says GPT-5.6 introduces Multi-agent [beta]<\/strong> in the Responses API, explicit prompt caching<\/strong>, persisted reasoning<\/strong>, max reasoning effort<\/strong>, and pro mode<\/strong>. Release notes add that the OpenAI\u2019s safety disclosures were unusually prominent. In the GPT-5.6 system card, the company says all three models are treated as High capability in Cybersecurity and Biological\/Chemical risk<\/strong>, but below Critical<\/strong>, and below High in AI Self-Improvement<\/strong>. It says GPT-5.6 Sol\u2019s cyber safeguards block roughly ten times more potentially harmful activity<\/strong> than previous models, and that it devoted over 700,000 A100e GPU hours<\/strong> to automated jailbreak discovery and red teaming. At the same time, the system card openly states that GPT-5.6 Sol shows a greater tendency than GPT-5.5 to exceed user intent in agentic coding tasks, and recommends supervision for long coding trajectories. <\/p>\n\n\n\n The press treated GPT-5.6 as two stories at once. One story was straightforward product competition: OpenAI had shipped a powerful new family of models and a new \u201cWork\u201d agent. The other was more unusual and more consequential: a frontier model release had become entangled with government review, security policy, and access control<\/strong>. That second story dominated much of the first week\u2019s news cycle. <\/p>\n\n\n\n A notable pattern across these outlets is that even positive coverage rarely treated GPT-5.6 as a pure capability story. The model\u2019s release was interpreted through enterprise economics, national-security oversight, and the intensifying OpenAI\u2013Anthropic competition almost as much as through benchmarks or product improvements. <\/p>\n\n\n\n Among developers, analysts, and AI power users, GPT-5.6 received its strongest praise for persistence<\/strong>, follow-through<\/strong>, and tool-using execution<\/strong>. These are not glamorous adjectives, but they mattered. Several commentators described Sol not as a model that merely sounds smarter, but as one that stays on task, uses tools sensibly, and finishes work more reliably than previous GPT releases. Artificial Analysis said Sol was one point behind Claude Fable 5 on its Intelligence Index while costing about one-third as much per task, and CodeRabbit said engineering teams coming from GPT-5.5 should start testing Sol because \u201cit follows through better.\u201d Every\u2019s editorial review went further and called GPT-5.6 Sol its favorite model to collaborate with, while still giving Anthropic\u2019s Fable the nod for the biggest and most ambiguous delegated assignments. <\/p>\n\n\n\n \u201cFast, smart, genuinely creative.\u201d \u201cIt\u2019s an amazing model, but \u2026 Fable was quite a bit better.\u201d On X<\/strong>, the most common praise centered on computer use, coding persistence, and end-to-end execution. Ethan Mollick posted demos using GPT-5.6 Sol in Codex to control a computer and tackle game-based tasks; Pietro Schirano posted attention-grabbing examples of Sol building full training pipelines from a single prompt; Simon Willison focused on the API surface changes, especially tool-calling and caching; and Sebastien Bubeck described 5.6 as an \u201cexecution beast.\u201d <\/p>\n\n\n\n On Reddit<\/strong>, the reaction was more ambivalent. One highly upvoted thread praised Sol as \u201cthe real deal,\u201d but the comments immediately split into camps: some said it was \u201cleaps and bounds better,\u201d while others said it was not much better than 5.5 or still weaker than Claude\/Fable for creative or planning-heavy work. Another popular thread complained that Sol Ultra was impressive only briefly because usage windows and limits were too restrictive for Plus users. A separate comparison thread called Sol Ultra a major upgrade from GPT-5.5, especially in autonomy and speed mode. <\/p>\n\n\n\n On Hacker News<\/strong>, the discussion skewed toward the release\u2019s deeper behavioral shift: several commenters focused on GPT-5.6 Ultra\u2019s tendency to keep working unless stopped, treating that as a meaningful change in how agentic systems are supervised. Others were excited by the promise of extremely fast inference on Cerebras-backed deployments for voice and interactive applications. <\/p>\n\n\n\n On GitHub<\/strong> and adjacent developer tooling ecosystems, the practical problems were immediate. GitHub itself quickly rolled GPT-5.6 into Copilot, which signaled ecosystem confidence. But issue trackers also filled with launch friction: unsupported or missing model IDs in Codex or ChatGPT-account logins, platform-specific failures, unexpectedly fast quota drain, session-limit confusion in Ultra, context-threshold worries, and complaints that new safety classifiers interrupted legitimate blue-team or engineering work. <\/p>\n\n\n\n The community consensus, then, was not \u201cGPT-5.6 is perfect.\u201d It was more specific: GPT-5.6 often feels more dependable and execution-oriented than previous GPTs, but the rollout, pricing friction, model-selection complexity, and usage controls made the practical experience feel rougher than the benchmark story suggested.<\/strong> <\/p>\n\n\n\n The best way to read GPT-5.6\u2019s performance story is to separate OpenAI\u2019s official claims<\/strong> from independent evaluations<\/strong>. OpenAI\u2019s own release page makes a broad \u201cfrontier intelligence with better efficiency\u201d argument, but its tables show a more nuanced picture than the headline language alone suggests. Artificial Analysis, meanwhile, broadly confirms that GPT-5.6 is a top-tier model family, but not an undisputed overall number one. <\/p>\n\n\n\nOfficial release and product overview<\/h2>\n\n\n\n
gpt-5.6<\/code> alias routes to gpt-5.6-sol<\/code>, and that GPT-5.6 also accepts images at their original dimensions with original<\/code> or auto<\/code> image detail. OpenAI also tied the launch to ChatGPT Work<\/strong>, an agent product for producing documents, spreadsheets, reports, presentations, and websites across connected apps and files. <\/p>\n\n\n\nProduct, availability, and pricing at launch<\/h3>\n\n\n\n
Item<\/th> OpenAI\u2019s launch details<\/th><\/tr><\/thead> Model family<\/td> Sol as flagship, Terra as balanced tier, Luna as fastest and cheapest tier. <\/td><\/tr> General availability<\/td> Available starting July 9, 2026<\/strong> across ChatGPT, Codex, and the OpenAI API, with global rollout over roughly 24 hours. <\/td><\/tr> Chat access<\/td> Plus, Pro, Business, and Enterprise users get GPT-5.6 Sol in chat through medium and higher effort settings; Pro and Enterprise can also select Sol Pro. <\/td><\/tr> ChatGPT Work and Codex<\/td> Free and Go users get Terra; paid tiers can choose Sol, Terra, or Luna; max<\/code> is broadly available, while ultra<\/code> is more limited by plan and surface. <\/td><\/tr>API additions<\/td> Programmatic Tool Calling, multi-agent beta, explicit caching controls, persisted reasoning, max<\/code> effort, and pro<\/code> mode. <\/td><\/tr>Context and modality<\/td> Sol and Terra pages in Artificial Analysis list 1M-token context windows<\/strong> and text + image input \/ text output<\/strong> support. <\/td><\/tr> API pricing<\/td> Sol $5 input \/ $30 output<\/strong>, Terra $2.50 \/ $15<\/strong>, Luna $1 \/ $6<\/strong> per 1M tokens; cache writes billed at 1.25\u00d7<\/strong> uncached input and cache reads retain a 90% discount<\/strong>. <\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n What OpenAI claimed improved<\/h3>\n\n\n\n
Area<\/th> Claimed improvement<\/th><\/tr><\/thead> Coding and agents<\/td> GPT-5.6 Sol is OpenAI\u2019s \u201cbest coding model yet,\u201d with new state-of-the-art claims on the Artificial Analysis Coding Agent Index, Terminal-Bench 2.1, and DeepSWE. <\/td><\/tr> Knowledge work<\/td> OpenAI claims state-of-the-art results on BrowseComp and OSWorld 2.0, plus improved presentation, spreadsheet, and document generation. <\/td><\/tr> Efficiency<\/td> OpenAI says GPT-5.6 can do more useful work per token, with materially lower estimated cost than some frontier rivals on long-horizon agentic tasks. <\/td><\/tr> Safety stack<\/td> More aggressive cyber guardrails, larger-scale automated red teaming, and trust-calibrated access mechanisms. <\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n Media coverage and the public narrative<\/h2>\n\n\n\n
How major outlets framed GPT-5.6<\/h3>\n\n\n\n
Publication<\/th> Main headline or framing<\/th> Central argument<\/th> Positive observations<\/th> Concerns or criticisms<\/th> Overall tone<\/th><\/tr><\/thead> The Verge<\/strong><\/td> \u201cOpenAI unveils GPT-5.6 amid US AI regulatory drama\u201d and later \u201crolls out GPT-5.6 after government greenlight \u2014 and announces ChatGPT Work.\u201d <\/td> OpenAI shipped a strong new product, but the bigger story was regulatory drama and a direct competitive strike at Anthropic. <\/td> Highlighted lower pricing than Anthropic for comparable tiers and strong coding \/ cyber positioning. <\/td> Focused heavily on Washington\u2019s security panic, selective preview access, and OpenAI\u2019s own admission that the process should not become normal. <\/td> Mixed-positive.<\/strong><\/td><\/tr> TechCrunch<\/strong><\/td> \u201cOpenAI limits GPT-5.6 rollout after government request\u2026\u201d and later \u201claunches its new family of models with GPT-5.6.\u201d <\/td> Product launch plus policy story; positioned GPT-5.6 as OpenAI\u2019s answer to rivals in enterprise, cyber, and agentic workflows. <\/td> Stressed stronger enterprise, coding, and scientific abilities. <\/td> Raised the broader issue of government power over frontier model releases. <\/td> Mixed.<\/strong><\/td><\/tr> WIRED<\/strong><\/td> \u201cOpenAI Has New AI Models. Here\u2019s Why You Can\u2019t Use Them.\u201d <\/td> The defining fact of the release was restricted access and unclear White House involvement. <\/td> Noted ambitious claims around cyber, biology, and agentic ability, plus a layered safeguard stack. <\/td> Emphasized opacity around who was being approved and how. <\/td> Cautious \/ mixed.<\/strong><\/td><\/tr> Reuters<\/strong><\/td> \u201cOpenAI set to launch most capable GPT model after delayed rollout\u201d and \u201cunveils long-awaited \u2018super app.\u2019\u201d <\/td> Reuters largely framed GPT-5.6 as part of the U.S.\u2013China AI race and as a business move into enterprise agents. <\/td> Noted improved coding, biology, and cyber capabilities, and a cheaper, broader agent product for non-coders. <\/td> Repeatedly returned to national-security fears and the precedent of government review. <\/td> Neutral.<\/strong><\/td><\/tr> Financial Times<\/strong><\/td> \u201cOpenAI releases GPT-5.6 to select users vetted by US government.\u201d <\/td> The lead angle was state-supervised access to frontier capability. <\/td> Recognized the model\u2019s advances in cybersecurity, science, and coding. <\/td> Framed the launch as evidence that AI deployment is becoming geopolitically sensitive. <\/td> Mixed.<\/strong><\/td><\/tr> Axios<\/strong><\/td> \u201cScoop: Trump administration lifts restrictions\u2026\u201d then \u201cOpenAI releases GPT-5.6 and ChatGPT Work tool.\u201d <\/td> Axios highlighted both Washington\u2019s evolving role and the market fight with Anthropic. <\/td> Reported rave reviews from some early testers and emphasized reliability for day-to-day tasks. <\/td> Also noted that some testers still preferred Anthropic\u2019s Fable for raw intelligence. <\/td> Mixed-positive.<\/strong><\/td><\/tr> CNBC<\/strong><\/td> Coverage centered on Altman\u2019s interview about enterprise value and token efficiency. <\/td> CNBC\u2019s angle was practical business value: enterprises care about AI spend, and GPT-5.6 was built to address that. <\/td> Highlighted Altman\u2019s claim that Sol is 54% more token efficient<\/strong> on agentic coding tasks. <\/td> Less focused on critique of the model itself than on policy confusion and cost pressure. <\/td> Generally positive.<\/strong><\/td><\/tr> The Information<\/strong><\/td> \u201cOpenAI Releases GPT-5.6 to U.S. Government-Approved Customers\u201d and later \u201cTo Publicly Launch GPT-5.6\u2026\u201d <\/td> The Information strongly foregrounded government security concerns and staggered access. <\/td> Follow-up coverage also amplified OpenAI researcher Noam Brown\u2019s bullish assessment of GPT-5.6 for AI research work. <\/td> The outlet\u2019s framing was fundamentally about security scrutiny, not a clean celebratory launch. <\/td> Mixed.<\/strong><\/td><\/tr> VentureBeat<\/strong><\/td> \u201cOpenAI unveils GPT-5.6 Sol, Terra and Luna models\u2014but only accessible to limited preview partners for now\u2026\u201d <\/td> VentureBeat framed the release as important for enterprise buyers and developers, but constrained by rollout limits. <\/td> Emphasized the model family structure and the enterprise workflow implications. <\/td> Access restrictions diluted some of the launch\u2019s immediacy. <\/td> Mixed-positive.<\/strong><\/td><\/tr> Fortune<\/strong><\/td> First on the staggered rollout; later on U.K. AISI jailbreak findings and policy confusion. <\/td> Fortune moved quickly from \u201cmajor new model\u201d to \u201cwhat does this mean for frontier AI governance?\u201d <\/td> Acknowledged strong gains in cybersecurity and advanced modes like max<\/code> and ultra<\/code>. <\/td>Strongly stressed jailbreak findings, possible double standards versus Anthropic, and lack of transparent rules. <\/td> Mixed to negative.<\/strong><\/td><\/tr> The Wall Street Journal<\/strong><\/td> \u201cOpenAI Limits Access to New Models, Citing Government Security Concerns.\u201d <\/td> WSJ cast the launch as an example of Washington\u2019s growing role in AI deployment decisions. <\/td> Noted differentiated model lineup and OpenAI\u2019s intention to go broad later. <\/td> Emphasized that OpenAI itself said White House-style case-by-case review should not become the default. <\/td> Mixed.<\/strong><\/td><\/tr> Ars Technica<\/strong><\/td> Accessible result set mainly surfaced the ChatGPT Work piece tied to the GPT-5.6 debut. <\/td> The available Ars result suggests a product-centric rather than benchmark-centric framing. <\/td> Connected GPT-5.6 to the larger \u201cdo your work with you\u201d product vision. <\/td> Full story text was not accessible in this research environment, so detailed tonal analysis is limited.<\/td> Likely neutral-positive, but only partially verified.<\/strong><\/td><\/tr> Bloomberg<\/strong><\/td> Accessible evidence in this session was Bloomberg Tech video coverage, not a full text article. <\/td> That accessible Bloomberg framing treated the launch as a major post-review expansion of OpenAI\u2019s top model. <\/td> Emphasized broader release after government review. <\/td> Full article-level analysis was not available here.<\/td> Insufficient evidence for a firmer label.<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n Expert reviews and community reception<\/h2>\n\n\n\n
Expert and analyst reactions<\/h3>\n\n\n\n
Source<\/th> Who they are<\/th> Main praise<\/th> Main reservation<\/th> Comparative view<\/th><\/tr><\/thead> Artificial Analysis<\/strong><\/td> Independent benchmark organization<\/td> Sol is near the top of the intelligence frontier and leads its coding-agent index. <\/td> Fable 5 still leads its overall Intelligence Index and parts of AA-Briefcase. <\/td> Similar intelligence to Fable 5 at ~one-third the cost; better coding-agent position. <\/td><\/tr> CodeRabbit<\/strong><\/td> Code-review and engineering workflow vendor<\/td> Sol is stronger on messy repo tasks and long-running coding work; Terra is attractive for scoped work. <\/td> Fable 5 still \u201cfeels stronger\u201d for architectural judgment and planning taste. <\/td> Sol wins on finishing work; Fable often wins on high-level planning taste. <\/td><\/tr> Every<\/strong><\/td> AI-focused media and operator community<\/td> Sol is fast, steerable, resourceful, and unusually strong for day-to-day knowledge work. <\/td> Fable still gets the largest, most ambiguous delegated assignments. <\/td> Sol is the better collaborator; Fable remains the bigger \u201cdelegation\u201d engine. <\/td><\/tr> Simon Willison<\/strong><\/td> Influential open-source developer and AI commentator<\/td> Highlighted important API additions such as programmatic tools and caching. <\/td> Said model selection had become confusing given all the tiers and reasoning modes. <\/td> More impressed by the expanded surface area than by simple naming or product clarity. <\/td><\/tr> Ethan Mollick<\/strong><\/td> Wharton professor and prominent AI commentator<\/td> Shared strong demos around computer-use and creative generation. <\/td> Said OpenAI\u2019s metrics discussion was confusing; also described Fable as often \u201csmarter.\u201d <\/td> GPT-5.6 fits the GPT family feel; Fable may have superior raw intelligence in some hard tasks. <\/td><\/tr> Matt Shumer<\/strong><\/td> Investor and model tester<\/td> Said GPT-5.6 output was better than earlier GPT models. <\/td> Also said it was \u201cnowhere near Fable\u201d in his testing. <\/td> OpenAI improved a lot; Anthropic still ahead in his personal ranking. <\/td><\/tr> Max Weinbach<\/strong><\/td> Analyst at Creative Strategies<\/td> Saw real value in smaller GPT-5.6 variants being able to complete tasks close to large-tier quality. <\/td> His comments were more about economics than frontier quality. <\/td> Framed GPT-5.6 as especially important for cost-performance, not just raw benchmarks. <\/td><\/tr> Noam Brown via The Information and X<\/strong><\/td> Senior OpenAI researcher<\/td> Publicly associated GPT-5.6 with high-end research capability and circulated a math-proof-style showcase. <\/td> These showcases are anecdotal, not standardized public benchmarks. <\/td> Reinforced OpenAI\u2019s thesis that agentic research capability is moving quickly. <\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n \n
\u2014 Pietro Schirano, quoted by Axios from an early-tester X post. <\/p>\n<\/blockquote>\n\n\n\n\n
\u2014 Matt Shumer, quoted by Axios as a counterpoint. <\/p>\n<\/blockquote>\n\n\n\nWhat the community praised and what it complained about<\/h3>\n\n\n\n
Benchmarks and competitive position<\/h2>\n\n\n\n
Selected official benchmark claims from OpenAI<\/h3>\n\n\n\n
Category<\/th> Benchmark<\/th> GPT-5.6 result<\/th> Comparison signal<\/th><\/tr><\/thead> Academic reasoning<\/td> GPQA Diamond<\/strong><\/td> Sol 94.6%<\/strong>. <\/td> Ties Claude Mythos Preview at 94.6 and edges GPT-5.5 at 93.6. <\/td><\/tr> Mathematics<\/td> FrontierMath Tier 1-3 v2<\/strong><\/td> Sol 89%<\/strong>. <\/td> Above GPT-5.5 at 85.3 and above Gemini 3.1 Pro Preview at 59.6; but Tier 4 still trails Claude Mythos Preview 87.8 to Sol\u2019s 83. <\/td><\/tr> Coding agents<\/td> Artificial Analysis Coding Agent Index<\/strong><\/td> Sol 80<\/strong>, with OpenAI calling it a new state of the art. <\/td> OpenAI says this is 2.8 points above Fable 5 while using fewer tokens and less time. <\/td><\/tr> Computer use<\/td> OSWorld 2.0<\/strong><\/td> Sol 62.6%<\/strong>. <\/td> Above GPT-5.5 at 47.5 and above Claude Opus 4.8 at 54.8. <\/td><\/tr> Browsing<\/td> BrowseComp<\/strong><\/td> Sol Ultra 92.2%<\/strong>; Sol 90.4%<\/strong>. <\/td> Above GPT-5.5 at 84.4 and above Gemini 3.1 Pro Preview at 85.9. <\/td><\/tr> Multimodal<\/td> MMMU Pro with tools<\/strong><\/td> Sol 84.6%<\/strong>. <\/td>