45 new flagship models.
One year-defining release.

The 2026 release timeline.

The Pareto frontier of 2026.

Each dot is one model. Up means better quality, right means more expensive. The top-left edge is the Pareto frontier — every dot under it is strictly dominated.

The first Haiku with adjustable reasoning, at $0.10/$0.50 below 100K input tokens — a tenth of Haiku 4.5's $1 input price. Artificial Analysis scores it 43, ahead of Gemini 3.8 Flash and GPT-6 Luna.

Mistral's ~1T-parameter flagship in hosted preview, with open weights promised for late October. Cheap at $0.68/$2.09, but early independent scoring puts it mid-table (Vals AI: 32nd of 44).

OpenAI's value pick, launched at DevDay. Artificial Analysis puts it one point below GPT-6 Astra at a fifth of Astra's price. OpenAI did not ship a GPT-6.1 Astra.

Near-Opus 5.5 on scoped work at half the price, and the default most people should use. Anthropic reports 70.6% on Terminal-Bench 4.0 (Opus 5.5: 66.4%) and 80.1% on OSWorld 2.1.

The new coding leader, and cheaper than the Opus 5 it replaces. Anthropic's table puts it at 89.9% on SWE-bench Pro, ahead of Fable 5.1 (81.2%) and Opus 5 (79.2%), at $4/$20.

A price cut more than a capability jump: $2/$10 against GPT-5.6 Sol's $4/$20 promotional rate, with some rows below its predecessor. GPT-6.1 Sol replaced it a week later.

OpenAI's budget tier at half the old price. It matches Claude Haiku 5.5 at $0.10/$0.50, but Artificial Analysis scores it lower (38 vs 43).

Grok 4.6's price with modest coding gains — Senior SWE-Bench pass@3 40.0% vs 38.9% (Snorkel AI). Independent SWE-Together runs found it using about twice Grok 4.6's tokens per task.

DeepSeek's new top API model: MIT-licensed, natively multimodal, 1M context. DeepSeek says it beats V4-Pro on coding and agents, and since September 14 it serves all V4-Pro requests. Its coding claims are not yet independently verified.

OpenAI's first GPT-6 model and its answer to Fable 5.1, two days later and at the same $10/$50. The new ceiling on computer use (72.6% OSWorld 2.0, ~47% less time per task than Sol), agentic coding (57.9% Terminal-Bench 4.0 vs Fable 5.1's 55.8%) and frontier math (97.6% FrontierMath Tier 4). Fable 5.1 keeps Humanity's Last Exam (65.0% vs 57.2%). No SWE-bench figure published.

Google's third Flash in six weeks, at 3.7 Flash's price. Google reports 90.8% on Terminal-Bench 2.1 (3.7 Flash: 81.6%), and Artificial Analysis measured about 302 output tokens per second. It is verbose, so judge it on cost per task.

A point release of Muse Spark at the same price, with more reliable tool calls and better first-attempt accuracy by Meta's account. No benchmark table was published for it.

The new frontier ceiling, and the rare upgrade that costs less than what it replaces. 52.6% on Terminal-Bench-Science 0.1 against Fable 5's 24.7%, and 55.8% on Terminal-Bench 4.0 against Opus 5's 52.3%. Base pricing is unchanged at $10/$50, but cache reads fell 75% to $0.25/1M — about 25% cheaper on typical workloads and up to 45% on agentic ones. Anthropic published no SWE-bench figure at launch.

Fable 5.1 with the safeguards lifted — the highest agentic-coding score Anthropic has published (60.9% Terminal-Bench 4.0), restricted to vetted US cybersecurity and life-sciences organisations. Identical price, context and cache discount to Fable 5.1.

The first natively multimodal GLM-5 — a 320B-A18B MoE with vision and video built in, MIT weights, a 1M context, and a fifteen-cent price. The cheapest serious multimodal model on the market.

Alibaba's Qwen4 architecture preview: a 125B MoE with just 6B active per token that posts SWE-bench Pro 62.5 and GPQA Diamond 91.7 — roughly GLM-5.2-class coding at a tenth of the running cost.

The clear GLM-5.2 upgrade at the exact same price — huge agentic gains (DeepSWE 46.2 → 66.9, Terminal-Bench 3.0 4.6 → 28.3) plus a first: 84.5% on CyberGym, narrowly ahead of Claude Mythos 5's 83.8% on security.

Google's best coding score per dollar — 80.8% SWE-bench Verified and GPQA Diamond 94.8% at half the price of 3.6 Flash, shipped three weeks after it. The catch: the price doubles to $1.50/$7.50 on January 1, 2027.

xAI's long-horizon agent play: finishes agent tasks in roughly half the turns of rivals, so real cost per completed task lands well under the sticker price. AA Intelligence Index 61 — 4th overall.

Meta's return to genuine open source — the first model from Meta Superintelligence Labs, a 30B dense Apache 2.0 release built for always-on agents. 76.0% SWE-bench Verified from a model that runs on a 24GB GPU.

Alibaba's 2.4T MoE flagship — a legitimate SWE-bench Pro upset over GPT-5.6 Sol (67.7 vs 64.6) and #2 globally for vision. The strongest Chinese multimodal model yet.

The best cheap agent engine of 2026: Terminal-Bench 82.7 at $0.14/M input, with MIT weights self-hostable in ~110 GB. Nothing touches its agentic capability per dollar.

The new premium coding default. Takes the SWE-bench Verified lead (~96%) at Opus 4.8's exact $5/$25 price, and gets within half a point of Fable 5 on agentic coding at half the cost.

The best Google model for agents: beats 3.5 Flash on every tested benchmark with ~17% fewer output tokens and cheaper output. Since followed by Gemini 3.7 Flash (August 13) and 3.8 Flash (September 2) at the same price.

The fastest model in Google's lineup at 350 tokens/sec, with real agentic chops at $0.30/M. GPT-5.6 Luna beats it on price-per-benchmark; Flash-Lite answers with speed and full multimodal input.

A cybersecurity fine-tune of 3.5 Flash that found 55 confirmed V8 issues vs 47 for the base model. Gated to governments and trusted partners via the CodeMender pilot — no public access or pricing.

The 1.6T MoE open-weights flagship: 80.6% SWE-bench Verified (self-reported) at an order of magnitude below closed-frontier pricing. Independent agentic scores land lower — but the value is still absurd.

The closest Chinese challenger to the frontier — #4 of all models on AA Intelligence Index, ahead of Opus 4.8. At 2.8T parameters, it's the largest open-weight release in history.

The budget disruptor of 2026. After the ~80% July 30 price cut, Luna posts scores within 2–4 points of Sol at 1/25th the output cost. Just don't trust it past 512K tokens.

xAI's coding-first model on the 1.5T V9 base. Wins SWE Marathon (29% pass@1, ahead of Opus 4.8 and Fable 5) and solves SWE tasks with ~4.2x fewer tokens than Opus 4.8.

The most agentic Sonnet yet and the new claude.ai default — 72.7% SWE-bench Verified (vs 62.3% for Sonnet 4.6) at the same price, with a $2/$10 promo through August 31.

OpenAI's flagship — Terminal-Bench 2.1 leader at 88.8% (91.9% in ultra mode with sub-agents) and the first frontier model to clear a US government review. Opus 5 still owns repo-level coding.

The sensible OpenAI default: within 1–4 points of Sol on core benchmarks at 60% lower output cost after the July 30 price cut. Makes GPT-5.5's rate card look obsolete.

The top open-weights coding model of mid-2026 — beats GPT-5.5 on SWE-bench Pro at roughly a sixth of the cost, MIT-licensed, with two reasoning-effort modes.

Open-weight 1T MoE coding specialist with only 32B active params — fast, cheap to serve, and ~30% more token-efficient than its predecessor.

The new global #1. 80.3% SWE-Bench Pro is an 11-point leap over Opus 4.8 (69.2%) — the biggest single-release jump of 2026. 1932 GDPval-AA, 1M context, native parallel subagents. Costs 2x Opus 4.8 ($10/$50), so reserve it for the hardest agentic and engineering work.

The same model as Fable 5 with safeguards lifted in high-risk areas — restricted to vetted partners for advanced cybersecurity and research. For everyone else, Fable 5 is identical at $10/$50 with standard safety controls.

The best-value premium model. 69.2% SWE-Bench Pro and 1890 Elo at $5/$25 — but Claude Fable 5 (80.3%) now leads the frontier at 2x the price, so Opus 4.8 is the smarter default for most premium work.

Alibaba's first closed-weight flagship — an agent-first model built for ~35-hour autonomous runs. Cracked the global top 5 on agentic evals at launch; superseded by Qwen 3.8 Max in August.

Google's I/O headliner: a Flash-tier model that beat Gemini 3.1 Pro on agentic benchmarks at ~278 tokens/sec. Superseded by 3.6 Flash two months later.

Europe's strongest open release of 2026. A clean middle option for teams that need a non-US model.

Best for agentic, computer-use, and Codex workflows. The right pick if your stack is already OpenAI-native.

77.6% SWE-bench Verified from a 128B dense open-weight model with vision — within ~2 points of Sonnet 4.6 at half the price, and actually self-hostable.

Was #1 on SWE-Bench Pro at 64.3% — now superseded by Opus 4.8 (69.2%) at the same price. Vision accuracy 98.5%, strong agentic recall. Still fully supported.

Anthropic's most powerful internal model. Found thousands of zero-days autonomously. Not released publicly.

600B open weights, a 5M-token context window, and Recursive Self-Improvement that re-checks its own reasoning mid-task. No major host has published verified per-token pricing yet — it joins our ranked catalog when one does.

Meta's first closed frontier model — GPT-5.5-tier intelligence with the broadest multimodal input (text, image, video, audio, PDF) at $1.25/$4.25. The surprise value play of the year.

OpenAI's best price/quality. Pair with Claude Opus for hybrid stacks — they're complementary, not competitive.

Google's Pro tier for 2026: strong research scores, native multimodality and Deep Think reasoning, with a 1M-token window at $2/$12.