Compare AI Models by Price Tier

Compare AI models by price-to-capability ratio. Three tiers, clear tradeoffs, no marketing fluff.

Model Tiers

Budget Tier: Under $1/1M Tokens

Price: $0.25–$1.00 per million input tokens
Best for: Prototyping, preprocessing, hobby projects, high-volume workflows

Models:

• Gemini 3 Flash — FREE input tokens, 1M context, 78% SWE-bench
• GPT-5 mini — $0.25/1M, cheapest OpenAI option
• Kimi K2.7 Code — $0.95/1M cache-miss input, cheaper routine Kimi coding API lane
• Claude Haiku 4.5 — $1.00/1M, fastest responses

Bottom line: Compare cost per successful task. Do not turn scores from different benchmark variants into one percentage of “frontier performance.”

Mid-Range Tier: $1–$3/1M Tokens

Price: $1.00–$3.00 per million input tokens
Best for: Production apps, daily coding, reliable reasoning

Models and lanes to compare:

• Claude Sonnet 5 — $2 / $10 through August 31, then $3 / $15, first Claude cost/performance test
• GLM-5.2 — $1.40 / $4.40 per 1M, July value pick to test for long-context coding
• Kimi K3 — $0.30 cache-hit / $3.00 cache-miss input, $15 output, Kimi flagship eval lane
• Kimi K2.7 Code — $0.19 cache-hit / $0.95 cache-miss input, $4.00 output, cheaper routine Kimi coding API lane
• MiniMax M3 — Standard PAYG from $0.60 / $2.40 per 1M, value coding-agent lane to test

Bottom line: This is the production spend band. Use it for daily work, then escalate only when the task proves it needs a premium model.

Premium Tier: $5+/1M Tokens

Price: $5.00+ per million input tokens Best for: Complex research, enterprise workloads, premium arbitration, and tasks where a better answer is worth the bill

Models:

• Claude Opus 4.8 — $5 / $25 per 1M, practical premium Claude baseline
• Claude Fable 5 — $10 / $50 per 1M, restored guarded escalation; revisit after Opus fails only if access, retention, and guardrail tests pass
• GPT-5.5 — $5 / $30 short-context standard, $10 / $45 long-context standard
• GPT-5.5 Pro — $30 / $180 short-context Pro lane

Preview watchlist: GPT-5.6 Sol, Terra, and Luna list prices from $1/$6 to $5/$30, but they are restricted to selected API organizations and Codex workspaces. Listed pricing does not make them active routes.

Bottom line: Do not promote premium as a default. Measure fallback/refusal behavior, cached-input economics, retention requirements, and cost per successful task.

Quick Selection Guide

Methodology

Pricing: Current official list prices, with subscriptions and API rates kept separate.

Benchmarks: Exact benchmark names and variants, with independent and vendor evidence labeled separately.

Use cases: Provider specifications plus explicit local-evaluation gaps; no implied hands-on result without an artifact.

See /verify/methodology/ for full verification standards.

For security-task caveats, see the LLM app-hacking field test. It separates generic model price/performance from cost per confirmed exploit on one deliberately vulnerable app.

Related Resources

• /compare/ — All comparisons (models + tools)
• /compare/models/glm-5.2-vs-kimi-k2.6/ — GLM-5.2 vs Kimi K2.6/K2.7
• /models/kimi-k3/ — Kimi K3 launch guide
• /models/claude-opus-4-8/ — Claude Opus 4.8 premium baseline guide
• /models/claude-sonnet-5/ — current Sonnet pricing and migration guide
• /models/gpt-5-6/ — GPT-5.6 family status, pricing, and benchmark evidence
• /models/ — Individual model deep-dives
• /value/free-stack/ — Free tier access guide
• /value/smart-spend/ — When to upgrade from free
• /lab/investigations/llm-app-hacking-field-test/ — Security field-test caveats

Last updated: July 18, 2026. Pricing, access, open-weight status, and benchmark positions are subject to change; verify the exact account and source before routing work.