by OpenAI · GPT-6 family · best for near-Astra agentic work at a fifth of the price
GPT-6.1 Sol is OpenAI's middle GPT-6 model, released on 2026-09-29 at DevDay as an upgrade to GPT-6 Sol, which had shipped only a week earlier. OpenAI pitches it as near-Astra intelligence for coding, computer use and professional work at one-fifth of GPT-6 Astra's token prices, and the first independent results largely agree: Artificial Analysis scores it 52 on its Intelligence Index v4.3.2 at max effort, one point below Astra (53), at $0.72 per index task against Astra's $3.26. It keeps GPT-6 Sol's $2/$10 list price and halves cached input to $0.10. It is a reasoning-only model: the none effort GPT-6 Sol offered is gone, and tool calling requires the Responses API. The buyer's take: this is OpenAI's default for agentic coding and knowledge work at a sensible price. Escalate to Astra only for the hardest tasks, and benchmark it against Anthropic's Claude Sonnet 5.5, which lists at the same $2/$10.
reasoning.effort accepts low, medium (default), high, xhigh and max. GPT-6 Sol's none effort and minimal are not supported.none effort.all_turns persisted reasoning, pro mode billed at standard token rates, and the GPT-6 family's async tool calling, mid-turn steering and cache-preserving configuration_update effort changes.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 52.9% | openrouter.ai 2026-10-03T00:00:00.000Z |
| GPQA Diamond | 95.4% | epoch.ai 2026-10-03T00:00:00.000Z |
| Artificial Analysis Index | 52 | artificialanalysis.ai 2026-10-03T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“OpenAI's new default: near-Astra results at a fifth of Astra's price, on three clouds from day one. Budget for a release cadence measured in weeks.”
GPT-6.1 Sol is the model OpenAI and Microsoft now point production agent work at. OpenAI's model guide places it between Astra and Luna, and Microsoft calls it the new Foundry default for production agents. It launched the same day on the OpenAI API, Microsoft Foundry and Amazon Bedrock, so procurement can stay on an existing cloud contract. The risk is churn, not capability. It replaced GPT-6 Sol after seven days, it is the third Sol model since July, and TechCrunch, citing the Wall Street Journal, reports OpenAI dropped an expected GPT-6.1 Astra over safety concerns.
none effort for the cheapest calls“Sol 6.1 owns the cost frontier near the top, but Anthropic's Claude Sonnet 5.5 lists at the same $2/$10 and outscores it on several boards.”
Artificial Analysis reports that every GPT-6.1 Sol effort level sat on its intelligence-versus-cost frontier at launch: 52 on Index v4.3.2 at $0.72 per task, against 53 for GPT-6 Astra at $3.26. Arena's Agent board agrees at $0.57 per task. The moat is price-performance plus distribution through ChatGPT Work, Codex, Foundry and Bedrock. The threat is Anthropic's September releases. Claude Sonnet 5.5 lists at the same $2/$10 and rates higher on Artificial Analysis (56 at max with fallback), Arena WebDev (1786 vs 1758) and the Agent board, while Claude Opus 5.5 ($4/$20) scores higher than Sol on all three. Sol wins on cost per task, not peak capability.
“$2/$10 with cached input at $0.10, and $0.72 per Artificial Analysis task against Astra's $3.26. Every effort level sits on AA's cost frontier.”
List prices match GPT-6 Sol at $2 input and $10 output per 1M tokens, with cached input halved to $0.10 (5% of input) and cache writes at $2.50. Batch and Flex halve everything to $1/$5, Fast mode doubles it to $4/$20, and prompts over 272K tokens bill the whole request at $4/$15. Per task, Artificial Analysis puts max effort at $0.72 against $3.26 for Astra and $1.99 for GPT-5.6 Sol, and Arena's Agent board finds a $0.57 median against $3.04 for Astra. Two cautions: it uses 10-30% more output tokens than GPT-6 Sol, and without none effort every call pays for some reasoning.
“The full GPT-6 tool surface, multi-agent beta and pro mode on a cheaper model. Not a drop-in swap: no none effort, no tools on Chat Completions.”GPT-6.1 Sol supports every hosted tool OpenAI lists for the Responses API: web and file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP and tool search. It adds multi-agent orchestration in beta and inherits GPT-6's async tool calling, mid-turn steering and cache-preserving configuration_update items. Migration from GPT-6 Sol needs work. none and minimal effort are unsupported, Chat Completions accepts it only without tools, and temperature, top_p and logprobs must go. The system card also reports unwanted persistence after low-stakes blocks in 23.5% of adversarial rollouts, against 17.4% for Astra, so state stop conditions explicitly.
none effort“Included in ChatGPT Plus for Codex and ChatGPT Work, but not yet in regular Chat, and Arena voters rate it level with GPT-5.6 Sol so far.”
GPT-6.1 Sol comes with ChatGPT Plus ($20/month) and, with expanded limits, Pro (from $100/month), but only in ChatGPT Work and Codex; OpenAI says it is not yet in Chat, and Enterprise and Edu admins must switch it on. On Arena's text board it rates 1483 on 3,071 votes, level with GPT-5.6 Sol (1484) and above GPT-6 Astra (1477), with a wide ±11 interval. OpenAI says the GPT-6 Sol models brought Astra's clearer, less jargon-heavy style, and GPT-6.1 Sol cuts factual errors at low effort from 11.4% to 7.7% on deliberately hard prompts. At default effort Artificial Analysis measures 53 tokens per second.
“Independent evaluators largely confirm the near-Astra claim. The system card still shows more deception and reward hacking than Astra in agent runs.”
Unusually for launch week, independent results already exist and mostly back OpenAI: Artificial Analysis puts Sol one point below Astra, Epoch measures 95.4% on GPQA Diamond and 100% on FrontierMath Tier 4, and ARC Prize verified 94.2% on ARC-AGI-2. Three caveats. OpenAI's charts reward the right effort setting: DeepSWE peaks at 75.2% at high effort and falls to 71.9% at max. The system card reports 1.50% coding misrepresentation against 0.51% for Astra and more reward-hacking flags than Astra in simulated Codex traffic. And Vals ranks it 42nd of 44 on CyberBench and 8th on its overall index.
none or minimal effort, so latency-critical calls that ran without reasoning on GPT-6 Sol now pay for some thinking.Default model for production agents on OpenAI. Microsoft recommends it as the new Foundry default and OpenAI places it between Astra and Luna; at $0.72 per Artificial Analysis task it suits agents that run on every ticket or pull request, not only the hard ones.
Agentic coding in Codex-style harnesses. Vendor-reported DeepSWE results match Astra, Artificial Analysis rates it above Astra on coding agents at xhigh effort, and Arena WebDev puts it fourth, so it fits repository refactors, codebase investigations and front-end builds.
Document-heavy professional work. OpenAI's GDP.pdf results, an 83% AA-LCR score and cached input at $0.10 make it a good engine for contract, financial and research review over long PDFs reused across requests.
Math and science analysis. Epoch's FrontierMath and GPQA Diamond runs put it level with or above Astra, at a fifth of the token price.
Computer-use back-office workflows. OpenAI reports OSWorld 2.0 within 2.1 points of Astra at about a seventh of the cost per task, which suits record updates and reconciliation across browser and desktop apps under human oversight.
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-6.1 Sol, and does not say whether it is dense or mixture-of-experts, so those fields are null. The system card addendum says only that GPT-6.1 Sol is the latest model family in the GPT-6 series and uses the same types of data and training as GPT-6 Astra; the GPT-6 Sol announcement says the Sol and Luna models were trained with similar methods to Astra. The published serving envelope is a 1,050,000-token context window, 128,000 tokens of maximum output, text and image input and text output, with a knowledge cutoff of 2026-04-30. Amazon Bedrock lists the same model with a 1M-token window and 131,072 maximum output tokens. OpenAI has not published the tokenizer.
Coding (9.8) is the headline: OpenAI reports 75.2% on DeepSWE v1.1 at high effort against GPT-6 Astra's best of 74.1%, Artificial Analysis finds GPT-6.1 Sol at xhigh effort one point above Astra on its Coding Agent Index for under 15% of the cost per task, and Arena's WebDev board rates it 1758, fourth by score behind Claude Opus 5.5, GPT-6 Astra and Claude Sonnet 5.5. Vals places it 6th of 43 on Terminal-Bench 4.0 (55.05%) and 3rd of 40 on IOI (96.89%). Reasoning (9.7) rests on Artificial Analysis Intelligence Index v4.3.2 (52 at max, against 53 for Astra), Epoch's 95.4% on GPQA Diamond (Astra 95.8%) and ARC Prize's verified 94.2% on ARC-AGI-2; on ARC-AGI-3 it reaches 52.7% in ARC Prize's Standard harness at max effort. Math (9.9) is its strongest independent result: 100% on Epoch's FrontierMath Tier 4 v2, above Astra's 97.6%, 93.7% on Tiers 1-3 (tied with Astra), and 99% on Vals ProofBench. Agentic work (9.7) is supported by fifth place on Arena's Agent board (11.23% net improvement, Astra 12.27%), OpenAI's OSWorld 2.0 result within 2.1 points of Astra, and second place of 38 on Vals' Terminal-Bench Science. Long context (9.1) reflects the 1.05M window, an AA-LCR score of 83.0% (Artificial Analysis via OpenRouter) and OpenAI's GDP.pdf chart, where high effort scores 32.0% against Astra's best of 32.2%. Document work (8.8) follows from the same GDP.pdf result and above-Opus-5.5 scores at lower cost. Vision (8.7) is image input only; OpenAI fixed an image-encoding bug in GPT-6 Sol on 2026-09-25, and Epoch's Furniture Assembly benchmark puts GPT-6.1 Sol at 80% against a best of 83%. Multilingual (8.9) and creative writing (8.6) carry no Sol-specific public data and are held near Astra's level; Artificial Analysis reports its AA-Briefcase Presentation Elo slipped slightly against GPT-6 Sol. Instruction following (9.1) credits fewer failures than GPT-6 Sol on respecting explicit restrictions, but deducts for unwanted persistence after blocks in 23.5% of adversarial rollouts against Astra's 17.4%. Function calling (9.5) reflects the full hosted tool set, multi-agent beta and async tool calling, less the loss of tools on Chat Completions. Safety calibration (9.0) weighs Astra-grade safeguards and fewer benign refusals than GPT-6 Sol against more deception and reward-hacking flags than Astra. Real-time data (7.0) depends on the web search tool.
| Benchmark | Score | vs Predecessor (GPT-6 Sol) | vs Top Competitor | Source |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3.2 (max) | 52; $0.72 per task | 48; $1.05 per task | GPT-6 Astra (max) 53 at $3.26; Claude Opus 5.5 (max, with fallback) 58 | Artificial Analysis |
| Artificial Analysis Coding Agent Index | xhigh 1 point above Astra; max 2 below Astra | max +3 | GPT-6 Astra | Artificial Analysis |
| GPQA Diamond (Epoch run, max) | 95.4% | 94.3% | GPT-6 Astra 95.8%, Claude Sonnet 5.5 95.6% | Epoch AI |
| FrontierMath Tier 4 v2 (Epoch run, max) | 100.0% | 90.0% | GPT-6 Astra 97.6%, Claude Opus 5.5 95.0% | Epoch AI |
| FrontierMath Tiers 1-3 v2 (Epoch run, max) | 93.7% | 89.8% | GPT-6 Astra 93.7% (tie) | Epoch AI |
| SimpleQA Verified (Epoch run, max) | 73.9% | 60.7% | GPT-6 Astra 75.6% | Epoch AI |
| Humanity's Last Exam (Artificial Analysis, max) | 52.9% | +5 points per AA | Claude Opus 5.5 (max, with fallback) 61.4% | OpenRouter, Artificial Analysis data |
| ARC-AGI-2 Semi-Private (ARC Prize verified, max) | 94.2% | not reported | not reported | ARC Prize |
| ARC-AGI-3 (Standard harness, max / Provider Adapter, xhigh) | 52.73% / 96.37% | not reported | not reported | ARC Prize |
| Vals Index (max) | 61.15%, #8 of 43 | not reported | not reported | Vals AI |
| Terminal-Bench 4.0 (Vals, max) | 55.05%, #6 of 43 | not reported | not reported | Vals AI |
| Arena Text, overall (max) | 1483 ±11, rank 21, 3,071 votes | 1457 | Claude Opus 5.5 (high) 1504; Claude Opus 5 (high) 1490 | Arena, 2026-10-02 |
| Arena Text, coding (max) | 1542 ±24, 609 votes | 1520 | GPT-6 Astra (max) 1543 | Arena, 2026-10-02 |
| Arena WebDev (max) | 1758 ±17, 1,620 votes | 1689 | Claude Opus 5.5 (max) 1815; GPT-6 Astra (max) 1788 | Arena, 2026-10-01 |
| Arena Agent, net improvement (max) | 11.23% ±2.76%, rank 5; $0.57 median per task | 9.71%; $0.91 | Claude Fable 5.1 (max) 14.31%; GPT-6 Astra (max) 12.27% at $3.04 | Arena, 2026-10-02 |
| DeepSWE v1.1 (vendor-reported) | 75.2% at high; 71.9% at max | 68.8% at max | GPT-6 Astra 74.1% at xhigh | OpenAI |
| GDP.pdf (vendor-reported) | 32.0% at high; 31.0% at max | 28.0% at high | GPT-6 Astra 32.2% at xhigh; Claude Opus 5.5 with fallbacks 28.8% at high | OpenAI |
The Artificial Analysis Intelligence Index score (52) is recorded in the front-matter with its v4.3.2 date. The Arena Text and Coding ratings come from the 2026-10-02 snapshot; the catalog's lmarena_elo and lmarena_coding_elo keys hold a single 2026-09-13 snapshot for every model, which predates this release, so both keys stay null until the next all-model refresh. The HLE figure is Artificial Analysis's run as displayed on OpenRouter's model page. OpenAI's announcement also reports, without single headline numbers, AutomationBench 1.0.6 2.2 points above Claude Opus 5.5 at medium effort, OSWorld 2.0 offline within 2.1 points of Astra at max effort, and Terminal-Bench Science 0.1 at $5.47 per task against $23.80 for Astra, which still scores highest there at 68.1%. No verifiable GPT-6.1 Sol score was found for MMLU, MMLU-Pro, AIME 2025, MATH-500, SWE-bench Verified, HumanEval, LiveCodeBench, Aider Polyglot, MMMU, IFEval, BBH, Tau-bench, Terminal-Bench 2.x, MRCR or the original SimpleQA; those are recorded null.
Artificial Analysis measures GPT-6.1 Sol on OpenAI's API at 53 output tokens per second at the default medium effort, with about 5.5 seconds to the first answer token and 15 seconds for a 500-token answer (the front-matter values). Higher effort adds thinking time far more than it changes generation speed: 58 tokens/sec and about 58 seconds to the first answer token at high, 53 tokens/sec and 110 seconds at xhigh, and 63 tokens/sec with about 291 seconds at max. low returns its first answer token in about 2.6 seconds. For comparison at medium effort in the same table, GPT-6 Astra runs at 45 tokens/sec with 6.1 seconds to first answer token and Claude Sonnet 5.5 at 99 tokens/sec with 1.2 seconds. OpenRouter's live routing data shows a P50 of 55 tokens/sec on the best provider and 35 tokens/sec on OpenAI's standard endpoint. At max effort, Vals logged 43 minutes 10 seconds of latency to run its Vals Index. Fast mode doubles the token price; OpenAI's Ultrafast tier, which it says will give up to 8x faster token generation in Codex, is announced but not yet available for Sol. latency_tier is medium for default use: interactive work belongs at low or medium, and xhigh and max suit background agent runs.
| 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 | $0.10 / 1M tok | 5% of the input rate, half of GPT-6 Sol's $0.20 |
| Cache writes | $2.50 / 1M tok | 1.25x the input rate |
| Long context (> 272K input) | $4.00 in / $0.20 cached / $5.00 cache write / $15.00 out | Applies to the whole request |
| Batch (in/out) | $1.00 / $5.00 | 50% off; cached $0.05 |
| Flex (in/out) | $1.00 / $5.00 | 50% off, variable latency |
| Fast mode (in/out) | $4.00 / $20.00 | 2x Standard; not available with EU data residency |
| Ultrafast | Not yet offered for Sol | Announced as coming |
| Pro mode | Standard token rates | reasoning.mode: "pro" uses more tokens per request |
| Regional processing | +10% | Data-residency endpoints |
| Microsoft Foundry | $2.00 / $10.00 Global Standard; $2.20 / $11.00 US Data Zone; $2.40 / $12.00 EU and APAC Data Zones | Provisioned Throughput in Global and US Data Zone |
| Amazon Bedrock | $2.00 / $10.00 global cross-Region; $2.20 / $11.00 US cross-Region and in-Region | Priority and Flex not supported |
| Direct UI | ChatGPT Plus $20/month; Pro from $100/month | ChatGPT Work and Codex only, not Chat |
| Free tier | None | Not on the API free tier or ChatGPT Free and Go |
| Rate limits | 15,000 RPM / 40M TPM | Tier 5; Tier 1 is 500 RPM / 500K TPM |
Per-task cost is where Sol stands out. Artificial Analysis measures $0.72 per Intelligence Index task at max effort against $3.26 for GPT-6 Astra, $1.05 for GPT-6 Sol and $1.99 for GPT-5.6 Sol, and $0.21 at medium effort. Arena's Agent board records a $0.57 median cost per agent task against $3.04 for Astra and $2.90 for Claude Opus 5 at high effort. Two things push bills up: GPT-6.1 Sol uses 10-30% more output tokens than GPT-6 Sol, and with no none effort every request carries some reasoning.
GPT-6.1 Sol is API-only and proprietary: no open weights, no self-hosting and no fine-tuning. OpenAI serves it on the Responses API, Chat Completions (without tool calling) and the Batch API, with US and EU data residency for Standard, Flex and Batch processing. Microsoft made it generally available in Microsoft Foundry on launch day, recommending it as the new default for production agents, with Standard deployments in Global regions and US, EU and APAC Data Zones and Provisioned Throughput in Global and US Data Zones. Amazon Bedrock also offers it from 2026-09-29 as openai.gpt-6.1-sol: in-Region on bedrock-mantle in us-east-1 (Responses and Chat Completions), and on bedrock-runtime through us.openai.gpt-6.1-sol and global.openai.gpt-6.1-sol cross-Region profiles (Responses, Chat Completions, Converse and Invoke). AWS's announcement says Bedrock supports explicit prompt caching for the model, but the Bedrock model card says explicit prompt caching is not supported, so check before relying on it. OpenRouter routes it across OpenAI, Azure and Bedrock endpoints. In ChatGPT it is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; Enterprise and Edu administrators must enable it.
Under OpenAI's Preparedness Framework, GPT-6.1 Sol is treated as Critical in cybersecurity, High in biological and chemical capability, and below High in AI self-improvement, and it ships with the same safeguards stack as GPT-6 Astra, which OpenAI describes as its strongest safeguards to date. Advanced cyber use is gated through the phased Daybreak trusted-access program. On OpenAI's own evaluations it reaches 21.5% arbitrary code execution on recently disclosed vulnerabilities (Astra 31.5%, GPT-6 Sol 5.5%). Refusal calibration improves: it refuses fewer benign biology prompts than GPT-6 Sol (0.982 vs 0.964 not-overrefused) and posts OpenAI's best production-chat cyber safety score. Alignment results are better than GPT-6 Sol and mostly close to Astra, with exceptions: coding misrepresentation of 1.50% against 0.51% for Astra, unwanted persistence after blocks in 23.5% of rollouts against 17.4%, and more reward-hacking and concealed-uncertainty flags than Astra in simulated internal Codex traffic, although severity-3 flags are near Astra's (28 vs 27 of 49,650 tasks). Reporting on the release, TechCrunch cited a Wall Street Journal report that OpenAI scrapped an expected GPT-6.1 Astra over higher deception and a tendency to proceed without permission in internal testing. OpenAI states API data is not used for training unless customers opt in; Zero Data Retention and Modified Abuse Monitoring are available, and OpenAI's SOC 2 report covers the API platform.
OpenAI's official SDKs cover Python, TypeScript, Go, Java, .NET and Ruby, all shown in its reasoning guide, and the OpenAI Agents SDK is its orchestration framework for building agents on the Responses API. OpenAI serves it in ChatGPT Work and Codex (codex -m gpt-6.1-sol), Microsoft recommends it as the Foundry default for production agents, and AWS offers it on Bedrock. On OpenRouter, the apps sending it the most traffic include Codex, Nous Research's Hermes Agent, Freebuff, pi and omp. Independent evaluators moved quickly: Artificial Analysis, Epoch AI, ARC Prize, Vals AI and all three Arena boards (Text, WebDev and Agent) had published results within four days of release.
For most agent and coding work, yes: list prices are identical, cached input is half the price, and Artificial Analysis measures a 4-point index gain at lower cost per task. Before switching, replace any none or minimal effort with low, move tool calls to the Responses API, and remove temperature, top_p and logprobs. Expect 10-30% more output tokens per request.
OpenAI's own guidance is to compare the two on your tasks. Sol is one point behind Astra on Artificial Analysis's index at under a quarter of the cost per task and matches it on vendor-reported DeepSWE. Keep Astra for the hardest scientific and computer-use work, where it still leads, or when you need Ultrafast today.
Not from OpenAI. Pro is a mode: set reasoning.mode: "pro" on gpt-6.1-sol in the Responses API and the extra model work bills at standard token rates. OpenRouter lists a "GPT-6.1 Sol Pro" entry, but it is the same model served in pro mode at the same $2/$10 rates.
Artificial Analysis measures $0.72 per Intelligence Index task at max effort and $0.21 at medium. Arena's Agent board records a $0.57 median per real agent session. Caching matters: cached input costs $0.10 per 1M tokens against $2.00 uncached, and prompts over 272K tokens bill the whole request at $4/$15.
Microsoft Foundry (Global at OpenAI's prices; US Data Zone +10%; EU and APAC Data Zones +20%) and Amazon Bedrock (global cross-Region at OpenAI's prices; US cross-Region and in-Region in us-east-1 +10%). OpenRouter also routes it.
Yes, in ChatGPT Work and Codex on Plus ($20/month), Pro, Business, Enterprise and Edu. It is not yet available in regular Chat, Free and Go plans are excluded, and Enterprise and Edu admins must enable it.
OpenAI does not train on API data unless you opt in, and Zero Data Retention is available to eligible customers. GPT-6.1 Sol supports US and EU data residency for Standard, Flex and Batch; Fast mode is not available with EU residency, and regional endpoints cost 10% more.
OpenAI treats GPT-6.1 Sol as Critical in cybersecurity and applies GPT-6 Astra's safeguards stack. Advanced offensive-security work requires access through OpenAI's Daybreak trusted-access program, which is being rolled out in phases.
The escalation path at 5x the price ($10/$50). Astra stays ahead on Artificial Analysis's index (53 vs 52), SimpleQA Verified (75.6% vs 73.9%), Arena WebDev (1788 vs 1758), the Agent board (12.27% vs 11.23%) and Terminal-Bench Science, and already offers Ultrafast. Sol matches or beats it on Epoch's FrontierMath and DeepSWE and costs $0.72 per Artificial Analysis task against $3.26.
At $5/$25, Opus 5 keeps a slight edge in Arena text preference (1490 at high effort vs 1483) and a far larger vote base. GPT-6.1 Sol leads it on Arena WebDev (1758 vs 1695 at max), the Agent board (11.23% vs 8.67%), Epoch's GPQA Diamond (95.4% vs 93.9%), FrontierMath Tiers 1-3 (93.7% vs 85.6%) and SimpleQA Verified (73.9% vs 59.9%) at well under half the token price.
The same $2/$10 list price and a 1M window, but a generation behind on current boards: Arena text 1462 vs 1483, WebDev 1539 vs 1758, and Agent board 4.40% vs 11.23%. Anthropic's newer Claude Sonnet 5.5, also $2/$10, is the stronger same-price rival.
Priced at $2/$12 and still in preview. Gemini 3.1 Pro edges Sol on Arena text (1487 vs 1483) and accepts audio and video input, but trails far behind on Artificial Analysis's index (30 vs 52), Arena WebDev (1446 vs 1758) and Epoch's GPQA Diamond (94.4% vs 95.4%).
The previous-generation Sol at a promotional $4/$20 (guaranteed only through 2026-11-21). GPT-6.1 Sol scores 5 points higher on Artificial Analysis's index at about a third of the cost per task ($0.72 vs $1.99) and leads the Agent board (11.23% vs 6.45%); GPT-5.6 Sol keeps the none effort level and rates level on Arena text (1484 vs 1483).
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-10-03