Kimi K3.1 ProMoonshot AI
Flagship reasoning model with always-on thinking, vision input and 1M context. $2 / $10. A Highspeed variant shipped the same day.
verify licenseIndex · updated 24 Oct 2026
Pick the right LLM, skip the noise.
A short, opinionated index of the models that matter right now, with prices, context windows and honest trade-offs. Test two, ship one.
Model radar
newest first · prices per MTok in/outMiniMax M3.1 FlashMiniMax
Multimodal agent model with tunable thinking depth and 1M context. From $0.10 / $0.50 on the cheapest host.
apiMercury 3Inception
Diffusion LLM that generates in parallel rather than token by token. 1M context, $0.10 / $0.35. Worth testing where latency dominates.
proprietaryDeepSeek V4.1-Flash 1017DeepSeek
Refresh that finishes tasks faster with shorter answers at the same API price. Native vision, 1M context, open weights (MIT, per trackers).
open weightsGPT-6 ProOpenAI
Listed by trackers with parallel agents. We could not confirm pricing or benchmarks from OpenAI itself yet.
verifyMuse Spark OpenMeta
816B total / 42B active MoE, natively multimodal, 1M context, $1.25 / $4.25 hosted. Weights under the Muse Model License (attribution; terms above $20M MaaS revenue).
open weightsNemotron 3.5 Bolt · Super VLNVIDIA
Open models at 74B (9B active) and 120B (12B active, vision-language). Details still thin; check the model cards before committing.
open weightsGemini 4 ArgonGoogle
Reported generally available after the Sep 30 gated announcement. Confirm pricing and limits in Google’s docs before planning around it.
verifyClaude Haiku 5.5Anthropic
Completes the 5.5 line next to Opus 5.5 ($4 / $20) and Sonnet 5.5 ($2 / $10). From $0.10 / $0.50, 1M context.
apiIf you test only three
full recommendations →Quality ceiling
GPT-6 Astra
Newest OpenAI frontier tier: 99.9% on ARC-AGI 3 and 100% on ExploitBench at $10 / $50. Use it where a wrong answer is expensive.
Agentic default
Claude Opus 5.5
1M context, 128K output and adaptive thinking at $4 / $20. Best fit for long-running coding and multi-step agents.
Volume & value
DeepSeek V4.1-Flash 1017
Vision, 1M context and open weights at roughly $0.15–$0.30 in / $0.60–$1.20 out. Pair it with a premium model above.
What changed this month
Sep – Oct 2026Claude 5.5 is complete
Opus 5.5 (Sep 22), Sonnet 5.5 (Sep 28) and Haiku 5.5 (Oct 7), all with 1M context. Fable 5.1 covers creative work; Mythos 5.1 stays gated. Anthropic also opened a waitlist research preview of Claude Duet, a background agent app that asks before acting.
GPT-6 gets tiers
Astra on top, Sol ($2 / $10) and Luna ($0.10 / $0.50) below, plus GPT-6.1 Sol for coding and computer use. A GPT-6 Pro with parallel agents was reported on Oct 16.
Open weights close in
Muse Spark Open (816B / 42B active), Mistral Large 4 (1.05T MoE), Kimi K3, Qwen3.8 and GLM-5.3 trade blows with closed models on coding and reasoning.
Gemini 4 Argon opens up
Gated to cyber defenders on Sep 30, now reported generally available. Gemini 3.8 Flash remains the proven production pick.
Cheap tiers got capable
$0.10 per MTok input now buys 1M context from Haiku 5.5, GPT-6 Luna, MiniMax M3.1 Flash and Mercury 3.
Diffusion enters the chat
Mercury 3 is the strongest diffusion LLM so far. Treat it as a latency specialist, not a general default.
Explore by need
Full model catalog
Searchable table of 70+ models with strengths, watch-outs and links.
Small team playbook
Two-tier architecture, eval habits and the traps that cost money.
Run it on your GPU
Find your VRAM, install LM Studio, chat locally in ten minutes.
Default stack
One premium model for quality, one cheap model for volume, an open model when data must stay home. Evaluate on your own prompts for a week before switching.