Compare LLMs side by side

Choose an LLM by filtering on your hardest constraint first — context window, compliance, or latency — then ranking the survivors by benchmarks that match your task and cost at your monthly volume. Relevant benchmarks differ by task: SWE-bench for coding, needle-in-haystack for RAG, MMMU for vision, τ-bench for agents. The LAXIMA LLM Picker automates this in 7 questions and returns a ranked shortlist with monthly cost estimates and the score breakdown behind each pick.

Yes — the compare panel at the top of this page puts any two models from the catalog head to head: input and output price, cached input, context window and max output, speed tier, throughput, time to first token, caching, vision, tool use, fine-tuning, open weights, EU and HIPAA flags, free tier, and a data-driven "choose this one if…" verdict for each. The best value in every row is highlighted. The comparison is computed in your browser from the same verified catalog the quiz uses, and the URL updates as you switch models, so you can share the exact matchup you are looking at.

Yes — click "Add a third model" in the compare panel and the table becomes a three-way comparison. Every metric row then highlights the leader across all three, and each model gets its own "choose this one if…" column reasoned against the other two rather than against a single rival. Three is the maximum, because a fourth column stops being readable on a laptop screen and the useful question is almost always a shortlist of three. Three-model comparisons are shareable too: the URL becomes ?compare=model-a-vs-model-b-vs-model-c.

Click any model name on this page and its full spec sheet opens in a pop-up — in the compare panel, in the comparison table headers, on the "best LLM for X" cards, and in the quiz results. The spec sheet shows full pricing (including cached input and batch discounts), context window and max output, throughput and time to first token, capability flags, deployment and compliance options, monthly cost at 10k / 100k / 1M calls, the trade-offs we know about, and which use cases we consider it the editorial pick for. Every spec sheet carries its own "verified" date.

As of July 2026, Claude Opus 5 is the strongest coding LLM — it tops SWE-bench Verified at 96.0% — but at $5 / $25 per 1M tokens most teams should reach for Claude Sonnet 5 ($3 / $15), which leads the price/performance tier on real-world coding and agentic benchmarks. GPT-5.6 Terra is a close third and got 20% cheaper on 2026-07-30 ($2 / $12) with strong tool use and fine-tuning support. For IDE-style fill-in-the-middle and autocomplete, Codestral 25.08 ($0.30 / $0.90 per 1M) is the best price/performance option with a 256K context window for large codebases.

Gemini 3.1 Pro leads for RAG in 2026 because of its 2M-token context window — the largest available from any frontier lab — and strong needle-in-haystack recall. Claude Sonnet 5 is a strong second with 1M context, prompt caching (up to 90% cost cut on repeat prompts), and excellent grounding. For on-prem or open-weights RAG deployments, GLM-5.2 and DeepSeek V4 are the leading self-hostable options, both with 1M-token context windows.

Claude Sonnet 5 is the best LLM for autonomous agents and tool use in 2026, leading τ-bench and real-world multi-step tool-calling benchmarks. Use Claude Opus 5 for the hardest agent loops where reasoning depth justifies the cost; use GPT-5.6 Terra when you need OpenAI's tool ecosystem (Code Interpreter, File Search, Built-in Retrieval).

GPT-5.6 Luna at $0.20 / $1.20 per 1M tokens is the cheapest credible frontier-lab LLM as of 2026-07-30, when OpenAI cut its price by 80% — it undercuts Gemini 3.1 Flash-Lite ($0.25 / $1.50), the previous holder, with Claude Haiku 4.5 ($1 / $5) a tier above. At high volume, prompt caching matters more than list price — it cuts repeated-prompt cost by 80–90% and is available on Claude, GPT-5, and Gemini. For open-weights / self-hosted workloads, DeepSeek V4 Flash is cheapest at $0.14 / $0.28 per 1M (just $0.0028 per 1M for cached input) when China data-routing is acceptable; GLM-5.2 (Z.ai) is a strong open-weights alternative you can self-host to avoid third-party data routing.

Gemini 3.1 Pro is the best LLM for maps and geospatial apps because Google's knowledge graph and Maps-adjacent training data give it a structural advantage in place names, routing, and travel reasoning. Gemini 3 Flash ($0.50 / $3) gives you the same geo advantage at production cost, and Gemini 3.6 Flash sits between them. Claude Sonnet 5 is a solid fallback when you need stronger tool use to chain Google Maps Platform APIs.

GLM-5.2 (Z.ai) is the best general-purpose open-weights LLM as of July 2026 — a usable 1M-token context window and the strongest open-weights coding scores (top Terminal-Bench 2.1 and SWE-bench Pro, #1 on Design Arena's code leaderboard). DeepSeek V4 (Flash at $0.14 / $0.28, Pro at $0.435 / $0.87 per 1M tokens) is the leading open-weights reasoning family when China data-routing is acceptable. Codestral 25.08 remains the best open-weights coding specialist with a 256K context. Note that Alibaba's flagship Qwen3.7-Max is now closed-weight, so it no longer counts as an open-weights option. Meta's Llama 4 family is excluded from this picker because it underperforms current open-weights leaders on independent benchmarks.

No — the picker is fully deterministic. It combines hard filters (modality, context window, compliance), soft scoring (benchmarks weighted by use case), and a curated editorial layer per use case. Every recommendation shows the full score breakdown. No inference calls are made, which means your answers stay on your device and the same inputs always produce the same ranking.

The catalog was last audited 2026-08-16 and every model's spec sheet carries its own "verified" date. New frontier models are added within a week of release; pricing and capability claims are reverified monthly against each provider's own pricing page, never an aggregator. If a model's verification date is more than 90 days old, treat pricing and capabilities as stale and confirm on the provider's page.

Because they still answer API calls, and if you are running one in production you need to compare it against its replacement before migrating. Models superseded by a newer sibling are marked Legacy rather than deleted: they stay selectable in the compare panel and keep full spec sheets, but they are excluded from the quiz results and the "best LLM for X" rankings, so the picker never steers a new build onto one. A model is only removed from the picker when its provider actually retires it. Where a provider has published a retirement date, the spec sheet shows it.

Cost estimates combine four inputs: (1) your selected monthly call volume, (2) typical token shape per use case — for example RAG averages 8k input / 600 output tokens per call, coding averages 3k / 800, (3) published provider pricing per 1M tokens, and (4) realistic prompt-caching assumptions for models that support it. Batch-eligible workloads apply the provider's 50% batch discount when you select "batch is fine" for latency. The spec-sheet tiles use a flat 2k input / 600 output shape with caching excluded so models stay comparable. Figures are ballpark, not contractual — always run a pilot before committing to production.

Editorial picks override the benchmark leader when a model has provider-specific strengths that generic benchmarks miss. Gemini 3.1 Pro wins for maps because of Google's knowledge graph; Codestral 25.08 wins for IDE autocomplete because of its fill-in-the-middle training. Editorial boosts are explicitly labeled in the result card with the reason shown, so you can always see when an opinion is being applied and decide if you agree.