LATEST MODELS · 28 OCT 2026
Understand Qwen3-Coder-Next’s open weights, coding-agent focus, hardware tradeoffs and how to compare specialist usefulness with general intelligence.
Read the full article →The leaders. The trade-offs. The model that fits your work.
Independent recommendations, connected to the evidence.
Our quality-first recommendation for a broad mix of demanding work. Its strong general benchmark position, web-coding result and agent ranking make it our most rounded starting point in this edition.
Why we chose it ↗Joint-highest displayed score in our Epoch snapshot. Confidence intervals overlap.
THE TRADE-OFFNot the lowest-cost option, and Gemini leads the text-preference snapshot. Best overall is our judgement—not a universal winner.
Switch benchmarks to understand where leaders agree—and where they differ.
ECI · bars start at zero
Bars start at zero within each source. Raw scores cannot be compared across sources. Model variants and snapshot dates differ; a missing entry does not mean a zero score. Country groups describe developers, not national capability.
Read our interpretation →Source snapshot: 25 Oct 2026. Displayed ECI estimates are rounded by the source. Ranges below the scores show 90% confidence intervals. Equal displayed scores retain source order; small gaps need not be statistically meaningful.
| Rank | Model | Developer | Score | Uncertainty |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | 169 | 165–174 |
| 2 | GPT-6 Astra | OpenAI | 169 | 165–175 |
| 3 | GPT-6.1 Sol | OpenAI | 168 | 165–174 |
| 4 | Claude Fable 5.1 | Anthropic | 166 | 163–171 |
| 5 | Gemini 4 Argon | Google DeepMind | 166 | 163–172 |
| 6 | Claude Sonnet 5.5 | Anthropic | 166 | 163–171 |
| 7 | Claude Opus 5 | Anthropic | 164 | 161–168 |
| 8 | GPT-6 Sol | OpenAI | 164 | 161–167 |
| 9 | Claude Fable 5 | Anthropic | 163 | 161–168 |
| 10 | GPT-5.5 Pro | OpenAI | 163 | 160–167 |
Original Epoch AI · General ECI source
Epoch AI data: CC BY 4.0. Selected and reformatted by AIstratica.
Source snapshot: 22 Oct 2026. Text leaderboard snapshot. ± values reproduce source uncertainty; overlapping intervals caution against declaring a clear winner. Evaluation settings are part of model names.
| Rank | Model | Developer | Score | Uncertainty |
|---|---|---|---|---|
| 1 | Gemini 4 Argon (High) | 1517 | ±7 | |
| 2 | Claude Opus 5.5 (High) | Anthropic | 1509 | ±7 |
| 3 | Muse Wonder (Max) | Meta | 1509 | ±8 |
| 4 | Claude Fable 5 (High) | Anthropic | 1503 | ±4 |
| 5 | Claude Fable 5.1 (Max) | Anthropic | 1503 | ±6 |
| 6 | Claude Opus 4.6 (High) | Anthropic | 1503 | ±3 |
| 7 | Claude Opus 4.7 (High) | Anthropic | 1501 | ±4 |
| 8 | Claude Opus 4.6 | Anthropic | 1497 | ±3 |
| 9 | Gemini 3.8 Flash (High) | 1496 | ±4 · preliminary | |
| 10 | Muse Spark 1.3 (Max) | Meta | 1496 | ±5 |
Original Arena · Overall text source
Source snapshot: 23 Oct 2026. Official evaluation leaderboard, text-only benchmarks, Overall tab, all model types. No confidence intervals were shown. This benchmark has different tasks and model coverage from Arena and Epoch.
| Rank | Model | Developer | Score | Uncertainty |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 (max) | Anthropic | 78.25 | Not supplied |
| 2 | Claude Sonnet 5.5 (max) | Anthropic | 75.56 | Not supplied |
| 3 | GPT-6.1 Sol (max) | OpenAI | 75.19 | Not supplied |
| 4 | GPT-6 Astra (max) | OpenAI | 74.3 | Not supplied |
| 5 | Gemini 4 Argon (high) | Google DeepMind | 72.52 | Not supplied |
| 6 | Doubao-Seed-2.1-Pro-0915 (high) | ByteDance | 71.8 | Not supplied |
| 7 | Qwen3.8-Max (xhigh) | Alibaba | 71.74 | Not supplied |
| 8 | Grok 4.7 (xhigh) | xAI | 71.48 | Not supplied |
| 9 | GLM-5.3 (max) | Zhipu AI | 71.45 | Not supplied |
| 10 | Gemini 3.8 Flash (high) | Google DeepMind | 71.01 | Not supplied |
Editorial picks reviewed 26 October 2026. Each recommendation has a specific scope and an alternative.
Arena’s web-development board puts Opus 5.5 (Max) first. Our choice for quality-first web-app work.
This is web-development evidence, not proof of leadership across every programming task.
Also consider: GPT-6 Astra.
The High variant leads the displayed Arena text snapshot. Our starting point for broad conversational work.
Preference scores do not measure factual reliability or source quality.
Also consider: Claude Opus 5.5.
The Max variant leads Arena’s displayed agent board. A strong candidate for delegated, multistep work.
Agent performance depends on tools and permissions. Keep approval over consequential actions.
Also consider: Claude Sonnet 5.5.
High ranks ninth in the text snapshot, at listed API rates of $0.75 input / $3.75 output per million tokens.
The result is preliminary; token prices are not total task cost. This pick is limited to the displayed top ten.
Also consider: Muse Spark 1.3.
For open-weight, research, image understanding and fastest-response winners, we need comparable task and release evidence. We do not award those titles from an overall text score. Explore the open-weight decision guide →
Pick a priority to see our starting recommendation.
We prioritise capability across more than one kind of evidence, then look for a specialist result that supports a practical use. Opus 5.5 combines a leading general ECI position with leadership on the retrieved web-development and agent boards. This supports a quality-first recommendation, without pretending it wins every task.
Our method is qualitative: general capability first, corroborating task evidence second, and cost, access and uncertainty as trade-offs. We do not average incompatible scores or assign invented percentages. GPT-6 Astra is the closest alternative in the displayed ECI scores; Gemini 4 Argon deserves attention when conversational preference is the main priority.
Epoch model profile · Arena web development · Arena agents · Full methodology
How strong is general capability across its benchmark collection? Use the confidence intervals: the leading rounded scores alone cannot settle a clear winner.
Which responses do participants prefer? A conversational favourite can differ from a benchmark leader. Preferences are not a guarantee of truth.
How do models perform on this evaluation suite? Read it alongside other evidence, while keeping its task coverage and evaluation settings separate.
Our retained OpenCompass snapshot places Opus first and Sonnet second. The latest table could not be independently re-read during this update, so its original 23 October date remains visible. The chart re-presents dated snapshots; it is not a live feed.
Check access in your country, current plan limits, data handling and usage rights. Review generated facts and code. Give agents only the permissions they need.
Use AI with confidence →Selected recent U.S. and Chinese developments. Announcement dates stay visible; a new release is not automatically a new winner.
Moonshot’s release notes confirm the rollout in Kimi Code, with a higher-speed variant and plan-dependent access to a one-million-token context window. The speed improvement is a provider claim.
Our take: A useful candidate for long coding workflows. A new checkpoint should earn a fresh comparison; it does not inherit K3’s earlier benchmark position.
Official announcement ↗Anthropic introduced an Opus 5.5-based assistant that monitors connected apps and works on delegated tasks. The research preview starts with a waitlist for U.S. Pro and Max subscribers.
Our take: The practical change is the connected workflow, not simply a higher model score. Access restrictions matter for readers outside the U.S.
Official announcement ↗Google announced an agent designed to connect business context, tools and work across applications. It frames the product around completing objectives rather than only answering questions.
Our take: For organisations, existing integrations and permission controls may matter more than a small benchmark lead. Check deployment eligibility before selecting it.
Official announcement ↗A model supplies the reasoning. An agent adds tools, connected data, memory and permission to act. The same model can perform differently in different agent systems.
Set a specific outcome and a clear success condition.
The model uses approved tools and checks intermediate results.
Inspect the output, approve consequential actions and retain a record.
Our shortlist is based on workflow fit. These are recommendations to investigate, not universal product rankings.
For quality-first coding and multistep work, start with Claude Opus 5.5 in a suitable agent environment, supported by the displayed Arena results. For long-codebase workflows, Kimi Code with K3.1 Pro is a recent Chinese contender to investigate; its provider claims need independent corroboration. Claude Duet and Google’s Gemini agent are product options whose access and integration fit must be checked separately.
Compare on five things: completion quality, tool compatibility, total task cost, human approval controls and recoverability after mistakes. A long context window or a high model score alone cannot establish the best agent.
Permissions and review checklist →Quality and preference do not point to exactly the same answer. Opus is our overall pick, while Gemini leads the retained text-preference snapshot. Choose around your work rather than a single league position.
This edition introduces task-based picks and a three-source chart. We have not claimed rank movements without comparable earlier snapshots. Future editions will explain changes when new evidence supports them.
Model spotlight: Kimi K3 → · Open vs closed →