LLM Models

Selection Shortcuts

GPT-6 and Opus 5.5 pointers were checked September 27, 2026; DeepSeek and GLM routing labels were reconciled October 2. October 3 adds Sol 6.1 identity, economics and access only; Sol 6’s benchmark rows retain their original revision and date. This is a selected comparison set, not a complete latest-release catalog. Anthropic’s model overview also lists Sonnet 5.5 and Fable 5.1; the Sonnet 5 and Fable 5 evidence below does not evaluate those successors. Other rows retain their own checked dates. Jump to model status, selected evidence, or history.

See Budget Tier, Mid-Range, Premium, and Smart Spend for workload-specific decisions.

September evaluation candidates

• GPT-6 Astra: evaluate Low for planning
• DeepSeek V4.1 Flash: API value and native vision
• MiMo V2.6: Flash value, Pro, plans and open weights
• September subscription and workload router

The older model summaries below retain dated benchmark evidence. Do not compare scores across model revisions or benchmark variants without checking the source date and harness.

New to model releases? Read open weights vs open source and how to read an RL training dashboard for the difference between downloadable artifacts, training signals and useful results.

Use this section to answer two separate questions: what is newest, and what can be used normally today. An announced model can be newer without being generally available.

Active Models

These selected lanes retain their source-specific review dates. “Active” describes availability, not newest-in-family status or an AIHackers accepted-result evaluation.

Claude Sonnet 5

The daily Claude production lane. Current provider pricing is $2/$10 input/output per million tokens; keep plan and route details tied to their source dates, then escalate to Opus 5.5 only when the task needs a stronger second pass.

Claude Opus 5.5

The current premium Claude route for architecture, consequential agentic coding, knowledge work, and final arbitration. The current provider price anchor is $4/$20 input/output per million tokens with $0.20 cache read; use the Opus 5.5 guide for access, benchmark provenance, and the extended independent review. The Opus 5 page remains a dated record in historical guides.

Kimi K3

Moonshot’s newest Kimi flagship: 2.8T vendor-stated parameters, 1M context, native vision, $0.30/$3/$15 cache-hit/input/output pricing, and Artificial Analysis Index 57. Official weights and serving materials are released under the Kimi K3 License; self-hosting cost and AIHackers CAR remain unverified.

Kimi K2.7 Code

Moonshot’s cheaper routine Kimi coding API lane, with base and HighSpeed model IDs, 256K context, multimodal input, required thinking mode, and $0.95/$4.00 cache-miss input/output pricing.

GPT-6.1 Sol, Sol 6, Astra and Luna

Use Sol 6.1 for complex-work evaluation where available: Standard API rates are $2/$10 ordinary input/output with $0.10 cached reads and $2.50 writes per million tokens. Sol 6 remains a separate model with $0.20 reads and dated benchmark evidence; its scores do not evaluate 6.1. Luna remains the focused-work candidate, and the Astra planning evaluation retains its own evidence. Use the pricing owner for context tiers, credits, plan access and source dates. Sol 6.1’s AIHackers controlled evaluation and CAR are not-run.

DeepSeek V4.1 Flash

Current direct-API value candidate with native vision and MIT weights. Use deepseek-flash; legacy Flash aliases now route to V4.1. The 0731 article preserves historical evidence.

Active supporting lanes

• GPT-5.5 and GPT-5.6 — retained for existing integrations and historical comparisons; they are not current default recommendations here.
• MiniMax M3 — 1M-context value coding-agent lane to test.
• Xiaomi MiMo V2.6 — Flash for value evaluation, Pro for harder agent work, and a separate 9B research checkpoint.
• Gemini 3 Flash — Gemini high-context value lane.
• Gemma 4 — local and open model family.

Guarded and restricted: Fable and Mythos

Claude Fable 5 and Mythos 5

Keep these models out of default routing for different reasons. Anthropic restored Fable globally on native Claude surfaces July 1, but included usage is plan-specific, cloud rollout is incomplete, and safeguards and retention still matter. Mythos remains approved-organization access.

• Fable 5: restored globally on native Claude surfaces; $10/$50 guarded escalation, not a default lane.
• Mythos 5: restored to approved US organizations; not a self-serve production option.
• Default premium lane: use Opus 5.5 after Sonnet 5; keep Fable guarded and Mythos access-specific.

Selected Evidence Table

Do not compare scores without matching the benchmark variant and harness. SWE-bench Verified, SWE-Bench Pro, Terminal-Bench, Artificial Analysis, vendor preference tests, and site-owned repository tests answer different questions. Use How to Read AI Benchmarks to turn those results into a shortlist before running your own tasks.

Historical Guides

These pages remain indexable for older integrations, pricing history, and search intent. They are not current recommendations. Historical comparison status does not by itself mean an API has been retired.

• DeepSeek V4 Flash 0731 — historical scores and MIT weights; the direct Flash aliases now serve V4.1 Flash.

• GLM-5.2 — historical July evidence: 1M context, 128K maximum output, $1.40/$4.40 pricing and Artificial Analysis Index 51; see Z.AI’s GLM-5.3 documentation for the newer revision. No GLM-5.3 result is inferred.

• Claude Opus 4.5 — historical Opus reference; use Opus 5.5.

• Claude Opus 5 — historical premium Claude record; use Opus 5.5 for current routing.

• Claude Opus 4.8 — prior premium baseline; use Opus 5.5 for current routing.

• Claude Sonnet 4.6 — prior daily-production baseline; use Sonnet 5 for current routing.

• Kimi K2.5 — historical and older free-access context; use K3 for latest-Kimi searches and K2.7 Code for cheaper routine coding.

• GLM-5.1 — prior GLM release; GLM-5.2 is also historical evidence, with GLM-5.3 documented separately by Z.AI.

• GLM 4.7 — legacy GLM guide.

• GPT-5.6 family — historical OpenAI family record; use the GPT-6 Sol/Luna pricing guide for current routing.

Older model names can still be correct when a tool genuinely exposes that version or a dated article describes the period. They should not be relabeled as current leaders.

Related links

• /compare/models/
• /value/smart-spend/
• /tools/
• /verify/methodology/

October 2 correction: DeepSeek/GLM routing, historical separation and evidence provenance. This is not a complete provider refresh; other model summaries retain their stated evidence dates, and known Anthropic successors still need separate evaluation.