The Best LLMs for AI Agents

Agent workloads punish models differently than chat does: a mediocre answer costs one turn, but a bad tool call, a hallucinated file path or a silent early exit costs the whole trajectory. The models below are ranked on the benchmarks that stress exactly that — Terminal-Bench, OSWorld, AutomationBench — plus the two things benchmarks miss: what the model costs per task (not per token), and how it fails.

How we picked

Same rules as every ranking on this site: sourced claims, dated pricing, generally available models only. Agent-specific additions: we weight multi-step benchmarks over single-turn ones, and we flag per-task economics wherever token consumption diverges from sticker price — which in this category is often. Gemini 4 Argon is limited to trusted defenders for now, so it isn’t ranked.

The short answer

All six, side by side

No single chart here on purpose: agent benchmarks run on different harnesses and versions (Terminal-Bench 2.1 vs 4.0, Codex vs Claude Code), and the same model can score very differently depending on who ran it. Charting them as one series would manufacture precision that doesn’t exist. The matrix labels each number’s provenance instead:

*Vendor-reported, on the vendor’s own harness. Cross-vendor numbers are not comparable — see the FAQ below.

The reasoning, pick by pick

1. Claude Sonnet 5.5 — the fleet default

Sonnet 5.5 (September 28) is the rare “best overall” that is also priced like a mid-tier model: $2/$10, the same list price as Sonnet 5, but with Terminal-Bench 4.0 at 70.6% (Sonnet 5 scored 10.3%), OSWorld 2.1 at 80.1%, and GDPval-AA v2.1 at 1844 against Opus 5.5’s 1846. Output is 30%+ faster and Anthropic estimates up to 30% lower cost per task because it uses fewer tokens; cache reads fell to $0.10 on October 7, which matters because cache reads dominate agent bills.

The limit is judgment. In Anthropic’s own testing Opus 5.5 is clearly stronger on complex, open-ended work, and Sonnet 5.5 works best on well-scoped tasks — which makes it the ideal subagent beneath an Opus orchestrator.

Strengths: near-Opus agent quality at half the price; 30%+ faster; 1M context; the first Sonnet with cyber safeguards and fallbacks. Watch out: weaker on ambiguous, long-running jobs; higher-risk cyber requests fall back to Sonnet 5; Anthropic later restated some numbers, so decimals differ between its pages.

2. Claude Opus 5.5 — the high-stakes agent

Opus 5.5 (September 22) leads Anthropic’s table on OSWorld 2.1 (81.8%), GDPval-AA v2.1 (1846), Terminal-Bench 4.0 (66.4%) and CursorBench 4.0 (57.8%), and ties Fable 5.1 for the lowest prompt-injection rate on Gray Swan’s benchmark. It ships with a screening classifier, an auditable open-source sandbox and code review for agent use — features that matter more once agents touch production systems. At $4/$20 it is 20% cheaper than Opus 5 on list price and about 40% cheaper to run on typical workloads.

It does not win everywhere: GPT-6 Astra beats it on Terminal-Bench-Science 0.1 (64.6% vs 58.7%) and AutomationBench (41.4% vs 40.0%) on Anthropic’s own table.

Strengths: computer-use leader; strong injection resistance and built-in agent safety tooling; 1M context and 128K output; fast mode at $8/$40. Watch out: thinking is always on; safeguards can reroute cyber, biology and frontier-LLM-development tasks; Anthropic cautions that benchmark margins are a less reliable guide at this level.

3. Claude Fable 5.1 — the autonomy ceiling

Fable 5.1 (September 1) is what you escalate to when a trajectory outgrows everything else. Ramp ran a 38-hour unattended machine-learning job in which it diagnosed a prior result as a label artifact and launched six parallel experiments, and Millennium used it to find a crash that nobody had explained in four to five years. AutomationBench rose to 31.4% from Fable 5’s 17.1%, and cache reads dropped 75% to $0.25, which Anthropic says cuts a highly agentic bill by up to 45%.

Budget carefully. $10/$50 is 2.5x Opus 5.5, and on Anthropic’s own table Opus 5.5 beats Fable 5.1 on every row it shows. Anthropic’s advice is to start on Opus and move to Fable only when your evals at higher effort still fall short. And classifiers can reroute cyber and biology tasks or score them zero — design your harness to tolerate the model swap.

Strengths: built for multi-day asynchronous runs; strong customer evidence on long-run reliability; retuned safeguards with 60% fewer cyber false positives. Watch out: 2.5x Opus 5.5’s price; classifier reroutes; scores lower than Opus 5.5 on Anthropic’s own table; Mythos 5.1, the less restricted sibling, is trusted-access only.

4. GPT-6.1 Sol — price-performance on business workflows

GPT-6.1 Sol (September 29) is OpenAI’s $2/$10 answer to Sonnet 5.5. OpenAI says it scores 2.2 points above Opus 5.5 on AutomationBench at about a third of the cost, comes within 2.1 points of the $10/$50 GPT-6 Astra on OSWorld 2.0 at about one-seventh of the cost per task, and more than doubles GPT-6 Sol on Terminal-Bench-Science 0.1 at $5.47 per task against $23.21 for Opus 5.5. It has a 1.05M context, cached input at $0.10, and the same three-tier ladder from Luna ($0.10/$0.50) to Astra for routing by difficulty.

The evidence is the weak spot: nearly everything is OpenAI’s own numbers, there is no Terminal-Bench 4.0 score, and Sol exists in two versions — new projects should target 6.1. GPT-6 Astra ($10/$50, same page) remains OpenAI’s top scorer on agentic science if cost isn’t the constraint.

Strengths: $2/$10 with 1.05M context; near-Astra results at a fifth of the price (OpenAI-reported); cached input at $0.10, matching Sonnet 5.5. Watch out: vendor-reported numbers; confusing ladder; text and image input only; Ultrafast for Codex was announced to follow.

5. Sakana Fugu — the orchestrator

Fugu is a different animal: itself a model trained to delegate, routing each task across a swappable pool of frontier and open models behind one OpenAI-compatible endpoint. Vendor-reported results put Fugu Ultra shoulder-to-shoulder with Fable 5 on hard engineering benchmarks — notable since no Anthropic model is in its pool. The trade-offs: orchestration tokens inflate bills on multi-agent runs, quality depends on routing, and it’s unavailable in the EU/EEA at launch.

Strengths: vendor resilience by architecture — the pool routes around any one provider’s outage; OpenAI-compatible drop-in; excels at long messy research and review tasks. Watch out: orchestration tokens make multi-agent runs cost more than one model call; parity claims are vendor-reported; EU/EEA unavailable at launch.

6. MiMo-V2.6-Pro — the open-weight pick

MiMo-V2.6-Pro (September 22) is a 1.02T MoE with 42B active, under an MIT license, that ingests text, images, video and audio with a 1M context and costs about $0.435/$0.87 via API. Xiaomi’s improvement came from six days of large-scale reinforcement learning, and it reports AutomationBench 53.1 (above the 50.3 it lists for Opus 5), Terminal-Bench 2.1 89.9 and Artificial Analysis Intelligence Index 46. It also takes video and audio input, which is useful for voice and video agents.

Keep the claims in proportion. Those are Xiaomi’s numbers on its own harness; on the harder Terminal-Bench 4.0 it scores 34.9 against 49.0 for Opus 5 in the same table. It is also a 1.02T-parameter model to host yourself, and older MiMo-V2.5 model IDs stop working on October 21 if you’re migrating.

Strengths: MIT license; omni-modal input; $0.435/$0.87 pricing; strong results for an open model on agent rows. Watch out: vendor-reported scores; trails the closed frontier on the hardest terminal work; self-hosting 1.02T parameters is an ops commitment.

Also considered

• GLM-5.3 and DeepSeek V4.1-Flash — MIT-licensed open options that lead on our coding ranking; both are text-first and strong on coding agents, with DeepSeek the cheapest at $0.15/$0.60 off-peak.
• Grok 4.7 — $2/$6 with a 500K context and Terminal-Bench 4.0 of 37.6%; a big step up from Grok 4.6 but behind the leaders.
• Mistral Large 4 — a 1.05T open-weight preview with strong cyber and legal results; weights are due at the end of October.

Choosing in practice

• Start with Sonnet 5.5 and measure per-task cost on your actual workload — effort settings move the bill more than the rate card does.
• Put Opus 5.5 on top. An Opus 5.5 orchestrator with Sonnet 5.5 (or Haiku 5.5) subagents fits how Anthropic positions the family.
• Escalate to Fable 5.1 only the trajectories that still fail on Opus 5.5 at higher effort.
• If you’re on OpenAI, benchmark GPT-6.1 Sol against Sonnet 5.5 on your own tasks — the vendor numbers can’t settle it.
• Keep approval gates on irreversible actions regardless of model. Calibration reduces silent failures; it doesn’t eliminate them.

Full specs for every model live in the models directory. Related: Best LLMs for Coding. This ranking reflects the field as of October 12, 2026.

Changelog

• 2026-10-12 — October update: Sonnet 5.5 replaces Sonnet 5; Opus 5.5 replaces Opus 5; Fable 5.1 replaces Fable 5; GPT-6.1 Sol replaces GPT-5.6; MiMo-V2.6-Pro replaces GLM-5.2 as the open-weight pick.
• 2026-07-29 — July update: Opus 5 replaces Opus 4.8 (OSWorld 2.0 + AutomationBench lead); GPT-5.6 GA replaces GPT-5.5 (Terminal-Bench 91.91%).
• 2026-07-02 — Initial ranking published: Sonnet 5 · Fable 5 · GPT-5.5 · Opus 4.8 · Sakana Fugu · GLM-5.2.

Frequently asked questions

Sources

• Anthropic — Introducing Claude Sonnet 5.5
• Anthropic — Introducing Claude Opus 5.5
• Anthropic — Claude Fable 5.1 and Claude Mythos 5.1
• OpenAI — Introducing GPT-6.1 Sol
• Sakana AI — Fugu product page
• Xiaomi MiMo — MiMo-V2.6
• Artificial Analysis — Intelligence Index & cost-per-task