01 Monthly rankings
LLM Rankings 2026
Claude Fable 5, GPT-5.6 Sol, Claude Opus 5, GPT-5.5, Gemini 3.1 Pro and the open-weight GLM-5.2 — compared on agentic benchmarks, prices and real scenarios.
TL;DR — which model for what (August 2026)
- Maximum capability: Claude Fable 5 — the first public Mythos-class model, 80.3% SWE-Bench Pro, 1M context; $10/$50 per million tokens.
- Terminal & agents: GPT-5.6 Sol — 88.8% Terminal-Bench 2.1, 91.9% in Ultra mode; Terra and Luna tiers cost far less.
- Everyday work: Claude Opus 5 — near Fable at half the price ($5/$25), effort dial low→xhigh.
- Your own perimeter: GLM-5.2, DeepSeek V4, Llama 4 — open weights; the gap to the frontier is measured in months.
Claude Fable 5
Anthropic · Mythos-class · Jun 9, 2026The first publicly available Mythos-class model — a tier above the entire Opus line. Record on SWE-Bench Pro: 80.3% versus 69.2% for Opus 4.8 and 58.6% for GPT-5.5. Output up to 128K tokens — roughly a hundred pages of code or documentation in one response. The release quirk: conservative safety classifiers route cybersecurity, biology, chemistry and distillation queries to Opus 4.8 — triggering in under 5% of sessions, with the substitution disclosed. Historical footnote: the release survived a government pause and returned to general access on July 1.
GPT-5.6 Sol
OpenAI · Jul 9, 2026The flagship of a three-tier family: Sol, Terra ($2.50/$15) and Luna ($1/$6). Best public result in terminal-agentic work — 88.8% on Terminal-Bench 2.1, with Sol Ultra mode reaching 91.9%, ahead of even the restricted Claude Mythos 5 (88.0%). The launch was unprecedented: at the U.S. government's request, access first went to vetted partners only. Strengths — command line, browsing, long tool chains.
Claude Opus 5
Anthropic · Jul 24, 2026The workhorse of the new line: close to Fable 5 on most tasks at half the price. New state-of-the-art on Frontier-Bench (43.3% versus Fable's own 33.7%) and GDPval-AA. The effort dial — low / medium / high / xhigh — trades cost against capability per request; at lower levels the model keeps most of its performance while burning far fewer tokens. Default model on the Max plan; per Anthropic, the most misuse-resistant Opus yet.
GPT-5.5
OpenAI · April 2026OpenAI's previous flagship remains a strong generalist: 82.7% on Terminal-Bench 2.0, 81.8% on CyberGym. Against the Mythos class it visibly trails on hard agentic code — 58.6% SWE-Bench Pro versus Fable 5's 80.3%. The Pro tier ($30/$180) adds accuracy on critical work. A sensible pick when your processes are already built around the OpenAI ecosystem.
Gemini 3.1 Pro
Google DeepMind · Feb 19, 2026The top of Google's line: 3.1 Pro for deep reasoning, Deep Think for the hardest problems, Flash 3.5 and 3.6 as the fast cheap layer for agent loops. Strengths — multimodality and ecosystem integration (Workspace, Vertex, Search). Fewer public results on cross-vendor agentic benchmarks than Anthropic or OpenAI, hence the dash in our table — missing data, not weakness.
MultimodalityGoogle ecosystemFew public agentic scoresGLM-5.2
Z.ai · Open SourceThe strongest open challenger to flagship level per independent comparisons this summer. Its open-source company: DeepSeek V4 (April 2026) competes on API price, Llama 4 Scout holds the context record — 10 million tokens. Open-weight developers themselves estimate the gap to the closed frontier in months, not generations. The usual price applies: your own GPU perimeter and engineering — in exchange, data never leaves your infrastructure. Full comparison in our open-vs-closed guide.
Data stays insideCost = hardware onlyNeeds own infrastructureComparison Table
| Model | Company | Release | Headline benchmark | Context | $/1M (in/out) |
|---|---|---|---|---|---|
| Claude Fable 5 | Anthropic | Jun 9, 26 | SWE-Bench Pro — 80.3% | 1M | 10 / 50 |
| GPT-5.6 Sol | OpenAI | Jul 9, 26 | Terminal-Bench 2.1 — 88.8% | 1.05M | 5 / 30 |
| Claude Opus 5 | Anthropic | Jul 24, 26 | Frontier-Bench — 43.3% (#1) | 1M | 5 / 25 |
| GPT-5.5 | OpenAI | Apr 26 | Terminal-Bench 2.0 — 82.7% | 1M | 5 / 30 |
| Gemini 3.1 Pro | Feb 19, 26 | — | n/a | Vertex rates | |
| GLM-5.2 | Z.ai · OSS | 2026 | open weights | — | self-hosted |
How to read the table. The "headline benchmark" column shows each model's best publicly confirmed result: SWE-Bench Pro — autonomous engineering in large repositories, Terminal-Bench — command-line work, Frontier-Bench — long office tasks. A dash means no public agentic data, not weakness. Context is a question of your documents: email fits in 128K, contracts want a million. Prices are per million tokens in/out; every vendor charges several times more for output.
"Headline benchmark" is each model's best publicly confirmed result; no single test has been publicly passed by all six as of August 2026. Sources: Anthropic and OpenAI announcements, Vellum, Vals AI, Artificial Analysis. Updated Aug 16, 2026.
Prices in Practice
A "$10 per million tokens" sticker says little until translated into tasks. Real volumes at August rates: a support chatbot — a typical dialogue runs 12 messages, about 4,000 tokens with knowledge-base context; a thousand dialogues a month is 4M input tokens: $20 on Claude Opus 5 or GPT-5.5, $40 on Fable 5, $4 on GPT-5.6 Luna. A 150-page contract analysis: ~110K input tokens — $0.55 per document on Opus 5 and $1.10 on Fable 5, versus an hour of a lawyer's time. Summarizing 500 emails daily: 2.5M tokens a month — from $2.50 on Luna to $25 on Fable.
The hidden line items people discover in month two. Output tokens cost several times more than input: $25–50 per million on flagships versus $5–10 in — generating long reports hurts the budget more than reading long documents. Reasoning bills by tokens actually "thought", and Fable 5's Adaptive Thinking never switches off — a hard question can cost several times a simple one at identical answer length. Re-sending the whole context with every message remains the chief budget-eater of chatbots; prompt caching, discounted up to 90% by flagship APIs, is the cure.
A budget anchor for teams up to 20 people: $100–300 a month in API fees covers assistants, document analysis and summarization without rationing. If the bill grows faster than the value — check caching and context length first, model choice second.
A Note on Switching Costs
One number worth knowing before committing to any vendor: in our reader survey, teams that kept prompts, evaluation sets and data pipelines separate from vendor SDKs reported switching models in days; teams that built directly against one API reported weeks and, in two cases, gave up entirely. The leaderboard above has changed order twice in twelve months — architecture that treats the model as a replaceable part is not paranoia, it is the base case.
How We Test
Since summer 2026 the familiar composite of MMLU, GPQA and HumanEval has been retired: classic tests are saturated, with flagships separated by fractions of a point. The ranking now rests on agentic benchmarks — SWE-Bench Pro (real engineering tasks in large repositories), Terminal-Bench 2.1 (autonomous command-line work), Frontier-Bench and GDPval-AA (long office work). Where the numbers come from: official vendor announcements cross-checked against independent trackers — Vals AI, Vellum, Artificial Analysis, Endor Labs. Discrepancies get flagged in the text: per Endor Labs, record-holder Fable 5 landed mid-table on safe vulnerability-fixing (59.8% FuncPass) despite outstanding official scores elsewhere — such gaps are not scandal but a normal property of young benchmarks.
Separately, a manual set of 14 working tasks that exist in no public benchmark: a 500-page contract with risks buried mid-document, a legacy code migration, a spreadsheet full of dirty data, a client email in three languages. These runs are more subjective but closer to reality — their results feed the long-form blog reviews, not the headline score.
Found a discrepancy with our numbers? Write to [email protected] — we re-check and correct publicly. That is the corrections policy at work.
Choosing by Scenario
The abstract "best AI" is a question without an answer. The right question is: best for what? Six situations from our practice:
200+ page contracts, expensive mistakes
Claude Fable 5 or Opus 5: the million-token context and the best results on long autonomous work. Opus 5 gives the same 1M at half the price; Fable is for when mistakes cost the most.
Support chatbot on a budget
Claude Sonnet 5 or GPT-5.6 Terra ($2.50/$15): on routine tickets customers will not notice the difference from flagships, and the API bill is several times lower. Bring in a flagship only for escalations.
Personal data, regulators, audits
Open-weight GLM-5.2, DeepSeek V4 or Llama 4 on your own GPUs — data never leaves the perimeter. Cloud is acceptable only on enterprise tiers with a contractual no-training clause.
Trends, breaking news, social
Grok 4.3 as the second model for real-time X data; writing and editing — Claude Opus 5 or GPT-5.6. No single model replaces this combination.
Meeting recordings, lectures, streams
The Gemini line: 3.1 Pro and Deep Think for deep analysis, Flash 3.5/3.6 for volume. Native video and the Google ecosystem remain their home turf.
Cheap mass operations: tagging, classification, digests
GPT-5.6 Luna at $1/$6, junior Claude tiers, or open models via inference providers. Price per token wins here, not benchmark percentages.
A separate tip for teams choosing their first model: skip the month-long bake-off. Take the three tasks that eat most of your time, run them through two or three APIs within a week, and count two things — how many edits each answer needs and what a thousand such requests cost. That homemade test predicts your experience better than anyone's leaderboard, including ours.