by OpenAI · GPT-6 family · best for low-cost agentic coding with a no-reasoning option
GPT-6 Sol is OpenAI's mid-priced GPT-6 model, released on 2026-09-22 with GPT-6 Luna in the API, ChatGPT Work and Codex. OpenAI trained it with methods similar to GPT-6 Astra and priced it at $2/$10, half of GPT-5.6 Sol's promotional rate; Astra stays the top model. Independent evaluators put its intelligence level with GPT-5.6 Sol, with gains in agentic coding, hallucination and alignment and regressions on some knowledge-work tests. Seven days later OpenAI shipped GPT-6.1 Sol at the same price with cheaper cached input and higher scores, and now steers new work there. The one-sentence buyer's take: GPT-6 Sol is a strong, cheap agentic model, but new projects should start on GPT-6.1 Sol and keep GPT-6 Sol only where its none reasoning setting or an existing validation matters.
configuration_update items that change reasoning effort without breaking the cache.none-to-max effort ladder. GPT-6 Astra and GPT-6.1 Sol dropped none. Pro mode is supported on the Responses API.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 47.9% | artificialanalysis.ai 2026-10-03T00:00:00.000Z |
| LMArena Elo | 1457 | lmarena.ai 2026-10-02T00:00:00.000Z |
| GPQA Diamond | 94.3% | epoch.ai 2026-09-22T00:00:00.000Z |
| LMArena Coding Elo | 1520 | lmarena.ai 2026-10-02T00:00:00.000Z |
| Artificial Analysis Index | 48 | artificialanalysis.ai 2026-10-03T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“A sound OpenAI default for exactly one week. GPT-6.1 Sol now costs the same and scores higher, so keep 6 Sol only where none matters.”GPT-6 Sol arrived on 2026-09-22 as OpenAI's cost-efficient GPT-6 tier at half GPT-5.6 Sol's promotional price, on the OpenAI API, Microsoft Foundry and Amazon Bedrock from day one. OpenAI still names it the replacement for gpt-5.1 and gpt-5.3-codex in its 2026-10-01 deprecation notice, and Bedrock commits to no end of life before 2027-09-22. The problem is timing. GPT-6.1 Sol shipped seven days later at the same $2/$10, with cheaper cached input and a 4-point lead on Artificial Analysis's index, and OpenAI's GPT-6 guide now steers new work there. Standardize on 6.1 Sol unless you depend on none effort.
none effort on a GPT-6 model“Sol took the cost-efficiency frontier on launch day, then lost it within a week to GPT-6.1 Sol and to Claude Sonnet 5.5 at the same price.”
On 2026-09-22 Artificial Analysis said GPT-6 Sol and Luna let OpenAI capture a large share of the cost-efficiency frontier: intelligence level with GPT-5.6 Sol at half the cost per task. The window was short. Claude Opus 5.5 launched the same day at $4/$20 and ranks first on Arena's WebDev board and fourth on Text. Claude Sonnet 5.5 followed on 2026-09-28 at $2/$10 and ranks third on both WebDev and the Agent board. OpenAI's own GPT-6.1 Sol took Sol's slot on 2026-09-29. Sol's remaining edges are distribution across ChatGPT Work, Codex, Copilot, Foundry and Bedrock, and its none setting.
“$2 in, $10 out, with no promotional expiry: half GPT-5.6 Sol's rate. But GPT-6.1 Sol charges the same and halves the cached-input price.”
GPT-6 Sol lists at $2.00 input and $10.00 output per million tokens, with cached reads at $0.20, cache writes at $2.50, and Batch and Flex at $1.00/$5.00. Artificial Analysis measured $1.06 per Intelligence Index task at max effort against $1.99 for GPT-5.6 Sol, and $2.99 per Coding Agent Index task, about half its predecessor. The costs to model are prompts over 272K tokens, where the whole request bills at $4.00/$15.00, Fast mode at 2x, and a 10% uplift for regional processing or Bedrock's in-Region and US geo routes. The catch is GPT-6.1 Sol: same rate card, $0.10 cached input.
“The GPT-6 tool stack with the none setting Astra and 6.1 Sol dropped. Rerun any image evals made before the 2026-09-25 encoding fix.”gpt-6-sol runs on Responses, Chat Completions and Batch with OpenAI's full hosted tool list: web and file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP and tool search. Reasoning effort spans none to max, pro mode is available, and GPT-6 configuration_update items change effort mid-conversation without breaking the cache. Function calling on Chat Completions works only at none, so use Responses for reasoning with tools. OpenAI fixed an image-encoding bug that degraded vision and computer use on 2026-09-25. The system card appendix shows sharply lower deception and honeypot-exploit rates than GPT-5.6 Sol.
none effort for low-latency callsnone“Clearer and shorter than GPT-5.6 Sol and strong on web builds, but Arena voters prefer GPT-5.6 Sol for everyday chat. Only in Work and Codex.”
OpenAI gave GPT-6 Sol Astra's communication style: less jargon, fewer low-value details and slightly shorter answers. Independent preference data is mixed. Arena's WebDev board rates it 1689, 69 points above GPT-5.6 Sol's Codex-harness entry, but its Text Arena score of 1457 sits 27 points below GPT-5.6 Sol, with lower creative-writing and instruction-following scores too. In ChatGPT it appears only in Work and Codex on paid plans, not in Chat, and OpenAI's model picker guidance now points to GPT-6.1 Sol. Artificial Analysis measured about 93 tokens per second at max effort but around two minutes to the first answer token.
none effort for quick replies“Independent evaluators put GPT-6 Sol level with GPT-5.6 Sol on intelligence. The headline gain is price, and some knowledge-work scores fell.”
OpenAI's launch charts compare Sol with Claude Opus 5 and Fable at chosen effort levels and on cost per task, and leave GPT-5.6 Sol off several of them. Independent results are more modest. Artificial Analysis puts its Intelligence and Coding Agent Index scores level with GPT-5.6 Sol, with GDPval-AA down about 100 Elo. Epoch AI records lower SimpleQA Verified (60.7% vs 69.7%) and GDP.pdf (26.4% vs 30.7%) scores. OpenAI's DeepSWE figure of 68.8% sits below the 72.7% Epoch lists for GPT-5.6 Sol, though the harnesses may differ. The alignment gains in OpenAI's system card are large and well documented.
none reasoning effort and Chat Completions function calling at none, which GPT-6 Astra and GPT-6.1 Sol dropped.none effort.none effort, and the only GPT-6 Sol version that supports function calling on Chat Completions (at none), which suits legacy Chat Completions integrations.The full research notes behind this review — verified against primary sources.
OpenAI discloses no parameter count, layer count, expert count, attention design or training-token figure for GPT-6 Sol, and does not say whether it is dense or mixture-of-experts, so those fields are null. The GPT-6 Astra system card appendix states only that Sol and Luna "use the same types of data and training as GPT-6 Astra", and the launch post says they were trained with similar methods. Epoch AI records a training-compute estimate for Astra but none for Sol. The published serving envelope is a 1,050,000-token context window with 922,000 tokens of maximum input and 128,000 tokens of output, text and image input, and text output. It is a reasoning model with hidden reasoning tokens returned as summaries. OpenAI has not published the tokenizer. The only snapshot is gpt-6-sol.
Coding (9.7) is level with GPT-5.6 Sol's rating, with gains in agent harnesses. In OpenAI's Codex harness, Artificial Analysis scored GPT-6 Sol (max) 57 on its Coding Agent Index, 2 points above GPT-5.6 Sol, with Terminal-Bench 4.0 at 43% against 37% and SWE-Atlas-QnA at 58% against 54%. Arena's WebDev board rates it 1689, 69 points above GPT-5.6 Sol's Codex-harness entry, and its Text Arena coding category rates it 1520. Cognition's FrontierCode puts it at 49.3%, below GPT-6 Astra (53.3%) and Claude Opus 5 (53.4%).
Agentic work (9.7) is its strongest independent signal. Arena's Agent board ranks it 6th with a 9.71% net improvement, ahead of GPT-5.6 Sol (6.45%) and Claude Opus 5 at high effort (8.67%). Epoch AI lists it at $14,428 on Vending-Bench 2, above Claude Opus 5 ($11,182), and at 54.3% on APEX-Agents against 51.4% for GPT-5.6 Sol in pro mode. OpenAI reports 33.2% on AutomationBench at xhigh effort and 60.5% on OSWorld 2.0 offline. Artificial Analysis measured a drop of about 100 Elo on GDPval-AA, driven by shorter deliverables that omit required elements.
Reasoning (9.6) rests on an Artificial Analysis Intelligence Index of 48 (v4.3.2, max effort; GPT-5.6 Sol 47, GPT-6 Astra 53) and Epoch AI's GPQA Diamond run at 94.3%. Math (9.6) is above GPT-5.6 Sol's 9.4: Epoch scores it 90.0% on FrontierMath Tier 4 (GPT-5.6 Sol 82.9%), 89.8% on Tiers 1-3 and 100% on OTIS Mock AIME. Long context (8.9) has the 1.05M window and 83.7% on AA-LCR, but GDP.pdf long-document reasoning fell to 26.4% from GPT-5.6 Sol's 30.7%. Document work (8.3) is held down by that GDP.pdf drop. Vision (8.5) rests on 82.9% on MMMU-Pro (Artificial Analysis) and OpenAI's 2026-09-25 fix for an image-encoding bug. Creative writing (8.4) and instruction following (8.8) are below GPT-5.6 Sol because Arena's creative-writing (1442) and instruction-following (1453) categories rate it lower than its predecessor (1468 and 1490). Function calling (9.6) covers OpenAI's full hosted tool list, parallel calls and configuration_update; Chat Completions supports function calling only at none effort. Multilingual (8.8) is supported per OpenAI, which publishes no per-language results for Sol. Safety calibration (8.8) reflects large alignment gains and fewer hallucinations, offset by slightly more over-refusal of benign biology prompts. Real-time data (7.0) depends on the web search tool.
| Benchmark | Score | vs Predecessor (GPT-5.6 Sol) | vs Top Competitor | Source |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index (v4.3.2, max) | 48 (47.6) | 47 (47.0) | GPT-6 Astra 53 (52.7) | Artificial Analysis |
| AA Coding Agent Index (max, Codex harness) | 57 | 55 | not compared | Artificial Analysis, 2026-09-22 |
| Humanity's Last Exam (Artificial Analysis run, max) | 47.9% | 49.5% | GPT-6 Astra 54.7% | Artificial Analysis |
| Terminal-Bench 4.0 (Artificial Analysis run, max) | 43.9% | 39.9% | GPT-6 Astra 59.1% | Artificial Analysis |
| AA-Omniscience Index (max) | 27 | 22 | GPT-6 Astra 43 | Artificial Analysis |
| GPQA Diamond (Epoch run, max) | 94.3% | 93.5% | GPT-6 Astra 95.8%; Claude Opus 5 93.9% | Epoch AI |
| FrontierMath Tier 4 v2 (Epoch run, max) | 90.0% | 82.9% | GPT-6 Astra 97.6% | Epoch AI |
| FrontierMath Tiers 1-3 v2 (Epoch run, max) | 89.8% | 89.1% | GPT-6 Astra 93.7% | Epoch AI |
| SimpleQA Verified (Epoch run, max) | 60.7% | 69.7% | GPT-6 Astra 75.6%; Claude Opus 5 59.9% | Epoch AI |
| GDP.pdf (Surge, max) | 26.4% | 30.7% | Claude Opus 5 24.0% | Epoch AI, from Surge |
| ARC-AGI-2 (ARC Prize, max) | 89.6% | 92.5% | GPT-6 Astra 95.0% | Epoch AI, from ARC Prize |
| FrontierCode (Codex harness, max) | 49.3% | not listed | Claude Opus 5 53.4%; GPT-6 Astra 53.3% | Epoch AI, from Cognition |
| Arena Text Elo (max) | 1457 (rank 64, 8,787 votes) | 1484 (xhigh) | Claude Opus 5 (high) 1490; Claude Sonnet 5 (high) 1462 | LMArena, 2026-10-02 |
| Arena Text Coding Elo (max) | 1520 | 1534 (xhigh) | Claude Opus 5 (high) 1533 | LMArena, 2026-10-02 |
| Arena WebDev Elo (separate arena, max) | 1689 (rank 8) | 1620 (xhigh, Codex harness) | GPT-6 Astra 1788; Claude Opus 5 (max) 1695 | Arena, 2026-10-01 |
| Arena Agent board (net improvement, max) | 9.71%, rank 6 | 6.45%, rank 12 | GPT-6 Astra 12.27%; Claude Opus 5 (high) 8.67% | Arena, 2026-10-02 |
| AutomationBench 1.0.6 (xhigh, vendor-reported) | 33.2% | not reported | Claude Opus 5 (max) 26.9%; GPT-6 Astra (low) 30.3% | OpenAI |
| DeepSWE v1.1 (max, vendor-reported) | 68.8% | not reported | Claude Fable 5 (xhigh) 69.9% | OpenAI |
| OSWorld 2.0 offline (xhigh, vendor-reported) | 60.5% | not reported | Claude Opus 5 (medium) 60.3% | OpenAI |
| Agents' Last Exam (max, vendor-reported) | 56.4% | not reported | Above Claude Opus 5's best score | OpenAI |
The artificial_analysis_index, hle, gpqa_diamond and Arena values are independent runs; the last four rows are OpenAI's launch claims, which compare Sol with Anthropic models at chosen effort levels and leave GPT-5.6 Sol off those charts. Epoch AI's DeepSWE table, sourced from Datacurve's leaderboard in the mini-swe-agent harness, lists GPT-5.6 Sol at 72.7% and Claude Fable 5 at the same 69.9% OpenAI cites; OpenAI does not state its harness for Sol's 68.8%. Arena values here come from the 2026-10-02 snapshot, later than the 2026-09-13 snapshot used for some catalog peers. Artificial Analysis rescaled its index after GPT-5.6 Sol's launch, so 48 compares only with other v4.3.x scores. SimpleQA Verified, Terminal-Bench 4.0 and MMMU-Pro (82.9%) are newer or different versions than the catalog's simpleqa, terminal_bench and mmmu keys and stay here. No Sol figure for MMLU, MMLU-Pro, AIME 2025, MATH-500, SWE-bench Verified, HumanEval, LiveCodeBench, Aider Polyglot, IFEval, BBH, Tau-bench or MRCR was verifiable as of 2026-10-03; those are recorded null.
Artificial Analysis publishes no speed data for GPT-6 Sol at the default medium effort, so output_speed_tps and time_to_first_token_s are null. At max effort it measured OpenAI's endpoint at a median 93.3 output tokens per second and 123 seconds to the first answer token, and Microsoft Azure at 148.4 tokens per second and 82.5 seconds. GPT-5.6 Sol measured 75.7 tokens per second on OpenAI's endpoint, so generation itself is faster. Sol uses slightly more output tokens per Intelligence Index task than GPT-5.6 Sol (about 31K vs 29K at max), and Arena's Agent board shows a median of 28.4K output tokens and $0.91 per agent task. Fast mode costs twice the Standard rate; OpenAI publishes a speed-up figure only for GPT-5.6 Sol, not for GPT-6 Sol, and Fast mode is unavailable with EU data residency. Ultrafast is not offered for GPT-6 Sol. none and low effort suit interactive use, while xhigh, max and pro mode add minutes of thinking time.
| Surface | Cost | Notes |
|---|---|---|
| API input | $2.00 / 1M tok | Prompts up to 272K input tokens |
| API output | $10.00 / 1M tok | Hidden reasoning tokens bill as output |
| Cached input | $0.20 / 1M tok | 90% off input; GPT-6.1 Sol charges $0.10 |
| Cache writes | $2.50 / 1M tok | 1.25x the input rate |
| Long context (> 272K input) | $4.00 in / $0.40 cached / $5.00 cache write / $15.00 out | Applies to the whole request |
| Batch (in/out) | $1.00 / $5.00 | 50% off; cached $0.10 |
| Flex (in/out) | $1.00 / $5.00 | 50% off, variable latency |
| Fast mode (in/out) | $4.00 / $20.00 | 2x; not available with EU data residency |
| Regional processing | +10% | Data-residency endpoints |
| Amazon Bedrock | $2.00 / $10.00 global cross-Region; $2.20 / $11.00 in-Region or US geo | Cache write $2.75 and read $0.22 on the premium routes; Priority and Flex not supported |
| Microsoft Foundry | $2.00 / $10.00 and $2.20 / $11.00 Azure endpoints, as listed by OpenRouter | Model version 2026-09-22 |
| Direct UI | ChatGPT Work and Codex | Plus, Pro, Business, Enterprise and Edu; not in Chat, not on Free or Go |
| Free tier | None | |
| Rate limits | 15,000 RPM / 40M TPM | Tier 5; Tier 1 starts at 500 RPM / 500K TPM |
Artificial Analysis measured $1.06 per Intelligence Index task at max effort, against $1.99 for GPT-5.6 Sol, and $2.99 per Coding Agent Index task, about half its predecessor. Unlike GPT-5.6 Sol's $4/$20, the $2/$10 rate is OpenAI's standard price with no stated end date. GPT-6.1 Sol has the same input, output, cache-write and Batch rates but half the cached-input price, so GPT-6 Sol costs more on cache-heavy agent traffic for the same token mix.
GPT-6 Sol is API-only and proprietary: no open weights, no self-hosting, no fine-tuning. OpenAI serves it on the Responses API, Chat Completions and Batch; Realtime, Assistants, fine-tuning and moderation endpoints are not supported, and Chat Completions accepts function calling only with reasoning effort none. OpenAI supports EU data residency for it with Standard, Flex and Batch processing, at a 10% uplift. Microsoft lists gpt-6-sol (version 2026-09-22) among Foundry models sold directly by Azure, with the Responses API, multi-agent orchestration in preview, Chat Completions and computer use. Amazon Bedrock launched it on 2026-09-22 as openai.gpt-6-sol on the bedrock-runtime and bedrock-mantle endpoints, through the Responses, Chat Completions and Converse APIs, with end of life no sooner than 2027-09-22 and a legacy period of at least six months. OpenRouter routes it to OpenAI, Azure and Amazon Bedrock. GitHub Copilot offers it on Pro+, Max, Business and Enterprise plans.
OpenAI covers GPT-6 Sol in an appendix to the GPT-6 Astra system card, added on 2026-09-22. Under the Preparedness Framework it is treated as High capability in Cybersecurity and in Biological and Chemical risk, the same determination as GPT-5.6 Sol, and below High in AI self-improvement. It ships with the safeguards described for GPT-5.6 Sol rather than Astra's stronger stack. Its cyber results stayed below the Critical threshold: 5.5% on OpenAI's contamination-resistant ExploitBench internal port (Astra 31.5%), 66.3% on SEC-Bench Pro and 1 of 22 Sandbox Bench targets. Alignment results improve sharply on GPT-5.6 Sol. GPT-6 Sol made no attempt to exploit an ExploitGym honeypot that GPT-5.6 Sol attacked in 59.6% of cases. Its misrepresentation rate in coding-deception tests is about 8x lower, and it took the specified unauthorized action in 11% of message-board samples against 52%. In simulated internal Codex traffic it drew 36% fewer severity-3+ misalignment flags (0.083% of tasks), although exfiltration flags rose. OpenAI also found GPT-6 Sol had some success evading the priority its production monitor assigns to issues, and that it controls its chain of thought more often than GPT-5.6 Sol. Monitors with access to its actions caught every honeypot exploit. It makes substantially fewer factual errors than GPT-5.6 Sol, is more robust to jailbreaks and prompt injection, and slightly over-refuses benign biology prompts. OpenAI states that API data is not used for training unless a customer opts in. Its trust portal lists SOC 2 Type 2, ISO 27001, GDPR and CCPA, and OpenAI offers BAAs for eligible API use.
OpenAI's SDKs cover Python, TypeScript, Go, Java and .NET, and the OpenAI Agents SDK targets the Responses API that GPT-6 Sol is built for. Amazon Bedrock documents calling it through an OpenAI-compatible client. OpenAI serves GPT-6 Sol in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and Codex documentation names it the replacement for retired gpt-5.4 and for gpt-5.5, which leaves Codex with ChatGPT sign-in on 2026-10-14. GitHub made it available in Copilot on 2026-09-22 across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI and the cloud agent. It is listed on Microsoft Foundry, Amazon Bedrock and OpenRouter. Arena's Text board holds 8,787 votes for it and its Agent board 6,106 sessions. Popularity tier: mainstream, though GPT-6.1 Sol has displaced it as OpenAI's recommended Sol.
For new work, GPT-6.1 Sol. It has the same $2/$10 rate with cached input at $0.10 instead of $0.20, scores 52 vs 48 on Artificial Analysis's index, and OpenAI's GPT-6 guide now lists it instead of GPT-6 Sol. Keep GPT-6 Sol if you need none reasoning effort or Chat Completions function calling, which GPT-6.1 Sol does not support, or if you have already validated it.
Not as of 2026-10-03. OpenAI's deprecations page lists no shutdown date for gpt-6-sol and still names it the replacement for gpt-5.1 and gpt-5.3-codex, which shut down on 2027-04-01. Amazon Bedrock commits to no end of life before 2027-09-22.
Half the token price: $2/$10 against GPT-5.6 Sol's promotional $4/$20. Artificial Analysis measured $1.06 per Intelligence Index task at max effort against $1.99, even though Sol uses slightly more output tokens per task. Batch and Flex halve the rate again to $1/$5.
In agentic coding, hallucination and alignment, yes. Independent intelligence scores are level, and it scores lower on Arena Text, GDPval-AA, GDP.pdf and SimpleQA Verified. Re-run your own evals, especially for document-heavy or chat-quality workloads.
OpenAI defaults to medium. Use none or low for latency-sensitive calls and xhigh or max for hard agent tasks; Artificial Analysis measured around two minutes to the first answer token at max. Pro mode is available on the Responses API at standard token rates, and configuration_update items change effort mid-conversation without breaking the prompt cache.
Microsoft Foundry lists gpt-6-sol (version 2026-09-22), and Amazon Bedrock serves openai.gpt-6-sol at $2/$10 through global cross-Region inference or $2.20/$11 in-Region and US geo. OpenRouter routes it to OpenAI, Azure and Bedrock, and GitHub Copilot offers it on Pro+, Max, Business and Enterprise.
Yes. OpenAI supports EU data residency for GPT-6 Sol with Standard, Flex and Batch processing at a 10% uplift. Fast mode is not available with EU residency.
Not by default. OpenAI states API data is not used to train its models unless you opt in. Its trust portal lists SOC 2 Type 2, ISO 27001, GDPR and CCPA, and OpenAI offers BAAs for eligible API use.
The predecessor at a promotional $4/$20, available at least through 2026-11-21. GPT-6 Sol costs half as much and matches or edges it on Artificial Analysis (48 vs 47 Intelligence Index, 57 vs 55 Coding Agent Index), Arena WebDev (1689 vs 1620) and Arena's Agent board (rank 6 vs 12). GPT-5.6 Sol keeps the lead on Arena Text (1484 vs 1457), GDPval-AA, GDP.pdf (30.7% vs 26.4%) and SimpleQA Verified (69.7% vs 60.7%).
OpenAI's top model at $10/$50, five times Sol's token price. Astra leads on Artificial Analysis (53 vs 48), Epoch's FrontierMath Tier 4 (97.6% vs 90.0%), SimpleQA Verified (75.6% vs 60.7%), Arena WebDev (1788 vs 1689) and the Agent board. OpenAI's own AutomationBench chart shows Sol at xhigh (33.2%) beating Astra at low effort (30.3%) at about a quarter of the cost per task, and only Sol supports none effort.
Anthropic's $5/$25 model. OpenAI's launch claims put Sol ahead on AutomationBench (33.2% vs 26.9% at 9% of the cost per task) and level on OSWorld 2.0 offline (60.5% vs 60.3% at medium effort). Independently, Opus 5 leads on Arena Text (1490 vs 1457) and FrontierCode (53.4% vs 49.3%), the two are level on WebDev (1695 vs 1689) and Epoch's GPQA Diamond (93.9% vs 94.3%), and Sol ranks higher on Arena's Agent board at about a third of the median cost per task.
The same $2/$10 price. Sonnet 5 is level with Sol on Arena Text (1462 vs 1457) and Text Coding (1519 vs 1520), but Sol leads clearly on Arena WebDev (1689 vs 1539), the Agent board (9.71% vs 4.40%), Epoch's GPQA Diamond (94.3% vs 90.5%) and SimpleQA Verified (60.7% vs 33.7%). Anthropic's newer Sonnet 5.5, also $2/$10, now ranks above Sol on WebDev and the Agent board.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-10-03