LLM API Pricing in 2026: Every Major Model Compared
Per-million-token prices for Claude, GPT, Gemini, DeepSeek, Mistral, and Grok — plus caching and batch discounts — verified against vendor pricing pages.
By Imtiaz RayhanUpdated 7 min read
On this page
Verified September 1, 2026, from vendor pricing pages only. Flagships (in/out per MTok): Claude Fable 5 $10/$50, Opus 5 $5/$25, GPT-5.6 Sol $4/$20 (promo), Gemini 3.1 Pro Preview $2/$12. Workhorses: Sonnet 5 $2/$10 (permanent), GPT-5.6 Terra $2/$12, Gemini 3.7 Flash $0.75/$3.75 (intro). DeepSeek went peak/off-peak. Caching ~90% off input; batch 50%.
Key takeaways
- Three durable price tiers exist across every vendor: frontier ($4–10 in / $20–50 out), workhorse ($0.75–3 in / $3.75–15 out), and budget ($0.10–1 in / $0.30–5 out) — match the tier to the task before comparing vendors. The workhorse rung got markedly cheaper this summer.
- Output costs 3–6x input everywhere, and reasoning/thinking tokens bill as output — long-output and heavy-reasoning workloads dominate most bills.
- The discount stack is the real price list: prompt-cache reads run ~0.1x input across vendors, batch APIs take 50% off, and they compose — cached batch input can cost ~5% of list.
- Open-weights models via hosts (DeepSeek V4, Kimi, Qwen, GLM on Together/Fireworks) undercut proprietary mid-tiers — at the cost of operating the integration yourself.
- Prices churn — this page is refreshed on a cadence (date stamped above), and every number here was read from a vendor pricing page, never aggregators.
All prices are USD per million tokens, standard tier, read directly from vendor pricing pages on September 1, 2026. Prices change; this page is maintained on a refresh cadence (the Updated date above is the source of truth), and numbers we couldn't verify on a vendor page are omitted, not estimated.
Anthropic (Claude)
| Model | Input | Output | Cache read | Batch (in/out) | Context |
|---|---|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | $1.00 | $5.00 / $25.00 | 1M |
| Claude Opus 5 | $5.00 | $25.00 | $0.50 | $2.50 / $12.50 | 1M |
| Claude Sonnet 5 | $2.00 | $10.00 | $0.20 | $1.00 / $5.00 | 1M |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | $0.50 / $2.50 | 200K |
Cache writes cost 1.25x input (5-minute TTL) or 2x (1-hour); cache reads are 0.1x input. Batch is 50% off and stacks with caching. Notably, the 1M-context models include the full window at standard per-token pricing — no long-context surcharge. Two changes since July: Claude Opus 5 (July 24, 2026) took over the Opus slot at the same $5/$25 as Opus 4.8, which is now marked legacy at unchanged pricing; and on August 10 Anthropic made Sonnet 5's introductory $2/$10 the permanent list price — the scheduled reversion to $3/$15 was cancelled. Opus 4.1 was retired on August 5. Opus 5 and 4.8 also offer a first-party fast mode (research preview) billed at $10/$50.
OpenAI (GPT)
| Model | Input | Cached input | Output | Batch | Context |
|---|---|---|---|---|---|
| GPT-5.6 Sol | $4.00 (promo through Nov 21, 2026) | $0.40 | $20.00 | 50% off | ~1M |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 | 50% off | ~1M |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | 50% off | ~1M |
| GPT-5.5 | $5.00 | $0.50 | $30.00 | 50% off | ~1M |
| GPT-5.5-pro | $30.00 | — | $180.00 | — | ~1M |
| GPT-5.4 | $2.50 | $0.25 | $15.00 | 50% off | ~1M |
| GPT-5.4-mini | $0.75 | $0.075 | $4.50 | 50% off | 400K |
| GPT-5.4-nano | $0.20 | $0.02 | $1.25 | 50% off | 400K |
The GPT-5.6 family (Sol / Terra / Luna, released July 9, 2026) is OpenAI's newest line, and it has been cut twice since launch (July 30 and August 21); Sol's $4/$20 is promotional pricing guaranteed at least through November 21, 2026. Cached input is 0.1x across the lineup (pro models excepted), and GPT-5.6 introduces Anthropic-style cache writes at 1.25x input. Flex tier matches batch pricing; the Priority tier was renamed Fast mode on July 30 — 2x standard on GPT-5.6, 2.5x on GPT-5.5/5.4 by the table. One caveat the headline windows hide: on GPT-5.5, 5.4, the pro models, and the 5.6 family, prompts above 272K input tokens bill at 2x input / 1.5x output for the whole request. The pro tiers ($30/$180) remain the premium-reasoning outliers of the whole market.
Google (Gemini API)
| Model | Input | Output | Cache read | Batch | Context |
|---|---|---|---|---|---|
| Gemini 3.1 Pro Preview | $2.00 (≤200K) / $4.00 (>200K) | $12.00 / $18.00 | $0.20 / $0.40 | 50% off | ~1M |
| Gemini 3.7 Flash | $0.75 (intro; $1.50 from Jan 1, 2027) | $3.75 (intro; $7.50 from Jan 1, 2027) | $0.075 | 50% off | ~1M |
| Gemini 3.5 Flash | $1.50 | $9.00 (incl. thinking) | $0.15 | 50% off | ~1M |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | $0.03 | 50% off | ~1M |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | $0.025 | 50% off | ~1M |
Google tiers long context: the Pro model's input price doubles (output 1.5x) beyond 200K of context, and cache storage bills separately per MTok-hour. The summer's news is at the Flash end — Gemini 3.6 Flash (GA July 21) and Gemini 3.7 Flash (GA August 13) both launched at an introductory $0.75/$3.75 that runs through December 31, 2026 before reverting to $1.50/$7.50, undercutting the still-listed 3.5 Flash. Gemini 3.5 Flash-Lite (GA July 21) is the designated replacement for 3.1 Flash-Lite, which has a May 7, 2027 shutdown date.
DeepSeek, Mistral, xAI
| Model | Input | Output | Notes |
|---|---|---|---|
| DeepSeek V4 Flash | $0.44 peak / $0.22 off-peak | $1.32 / $0.66 | Cache hits $0.014 / $0.007; 1M context, 384K max output |
| DeepSeek V4 Pro | $1.32 peak / $0.66 off-peak | $3.96 / $1.98 | Cache hits $0.044 / $0.022; 1M context, 384K max output |
| Mistral Medium 3.5 | $1.50 | $7.50 | Open weights (Modified MIT); batch 50% |
| Mistral Large 3 | $0.50 | $1.50 | Open-weights flagship (not a typo: priced below Medium 3.5) |
| Mistral Small 4 | $0.15 | $0.60 | Budget tier; 256K context |
| Grok 4.6 (xAI) | $2.00 (<200K) / $4.00 (≥200K) | $6.00 / $12.00 | Newest flagship (August 2026); 500K context |
| Grok 4.3 (xAI) | $1.25 (<200K) / $2.50 (≥200K) | $2.50 / $5.00 | 1M context |
DeepSeek's August change is the biggest move on this page: peak/off-peak pricing took effect August 16, 2026 — off-peak rates are half of peak, with peak defined as 01:00–04:00 and 06:00–10:00 UTC on weekdays — and list prices rose with it (V4 Flash was $0.14/$0.28 flat in our July snapshot). Off-peak V4 Flash now sits inside the Western budget band rather than under it; the 1M window and 384K max output survive, and the legacy deepseek-chat/deepseek-reasoner aliases (slated for retirement July 24, 2026) no longer appear in DeepSeek's model list. xAI bills like Google: a prompt that reaches 200K tokens pays the higher rate on every token in the request.
Open-weights via hosts
Reference serverless prices (host pricing, not vendor MSRP), September 1, 2026 — Together AI / Fireworks:
| Model | Together (in/out) | Fireworks (in/out) |
|---|---|---|
| DeepSeek V4 Pro (0813) | $1.32 / $3.96 | $1.32 / $3.96 |
| Kimi K3 | $3.00 / $15.00 | $3.00 / $15.00 |
| Kimi K2.6 | — | $0.95 / $4.00 |
| Qwen 3.7 Plus | $0.32 / $1.28 | $0.40 / $1.60 |
| GLM-5.3 | $1.40 / $4.40 | $1.40 / $4.40 |
| GPT-OSS-120B | $0.15 / $0.60 | $0.15 / $0.60 |
Hosted open-weights sit squarely inside the proprietary mid-tier price band — and the newest ones reach its ceiling: Kimi K3 at $3/$15 matches Claude Sonnet 5's list price on both hosts, so open weights no longer guarantee a discount. That band is still the lever that keeps the whole market's pricing honest, and the input to any self-host decision.
Reading the table like an engineer
Three structural facts matter more than any single cell. Output dominates: at 3–6x input everywhere — and with reasoning thinking-tokens billed as output — verbose responses and deep deliberation drive bills more than prompt size. The discount stack is enormous: prompt-cache reads at ~0.1x input plus batch at 50% compose to ~95% off for cacheable offline work — engineering for the stack beats switching vendors. Tiers beat brands: every vendor offers frontier/workhorse/budget rungs; matching the tier to the task and measuring cost per completed task (not per token) is where the money actually is — the full playbook is LLM Cost and Latency Engineering.
Continue exploring
- LLM Context Windows Compared (2026) — Context windows and max output tokens across Claude, GPT, Gemini, DeepSeek, and Grok — the million-token era, what it costs, and what fits in practice.
Frequently asked questions
- What's the cheapest good LLM API in 2026?
- Depends on the floor you need. For mechanical work, GPT-5.6 Luna ($0.20/$1.20), GPT-5.4-nano ($0.20/$1.25), Mistral Small 4 ($0.15/$0.60), and Gemini 3.5 Flash-Lite ($0.30/$2.50) define the budget tier; Claude Haiku 4.5 ($1/$5) is the premium-budget pick. For workhorse coding/agent use, Sonnet 5 ($2/$10), GPT-5.6 Terra ($2/$12), and Gemini 3.7 Flash ($0.75/$3.75 introductory) now sit well under the old $3/$15 line — capability per dollar there is task-dependent: benchmark on your work.
- How do prompt caching and batch discounts actually combine?
- Multiplicatively, on most platforms. Example on Claude: batch halves list price, and cache reads are 0.1x input — Anthropic's batch+cache pricing means a cached input token in a batch job costs ~5% of the standard input rate. Structuring for cacheable prefixes and routing offline work to batch APIs are the two highest-ROI cost moves before touching model choice.
- Do 1M-token context windows cost extra?
- On Anthropic, no — Fable 5, Opus 5, and Sonnet 5 include the full 1M window at standard per-token pricing (a 900k-token request bills at the same rate as a 9k one). The others tier: OpenAI bills GPT-5.6/5.5/5.4 prompts above 272K input tokens at 2x input and 1.5x output, Gemini 3.1 Pro doubles input beyond 200K, and xAI charges double on Grok prompts of 200K+. Either way you pay for the tokens you send — big windows make big bills possible, not free.
- Why do you only cite vendor pricing pages?
- Because third-party price aggregators drift stale within weeks and propagate each other's errors. Every figure on this page was fetched from the provider's own pricing or docs page on the date stamped above; anything we couldn't verify that way is omitted rather than guessed.
Filed under
pricing · llm-api · costs · comparison · data
Related
- LLM Context Windows Compared (2026)Context windows and max output tokens across Claude, GPT, Gemini, DeepSeek, and Grok — the million-token era, what it costs, and what fits in practice.
- Calling Any Model: Unified LLM Gateways & SDKs in 2026Why teams put a unified layer in front of LLM providers — and how LiteLLM, OpenRouter, and the Vercel AI SDK compare for fallback and cost control.
- Batch InferenceBatch inference processes many LLM requests asynchronously instead of one-at-a-time interactively — typically at ~50% discount via provider batch APIs.
- Prompt CachingPrompt caching reuses the computed state of a repeated prompt prefix across requests — dramatically cutting cost and time-to-first-token for stable context.
- LiteLLM vs OpenRouter: One API for Every Model (2026)LiteLLM vs OpenRouter compared — self-hosted gateway library vs hosted model marketplace. Keys, billing, control, and which unified LLM layer fits.
- AI WrapperAn AI wrapper is a product whose core function is a prompt and interface around a third-party model API, such as Claude or GPT, rather than its own model.
- LLM Cost and Latency Engineering: Caching, Right-Sizing, and p95 BudgetsA practical playbook for cutting LLM cost and tail latency — caching, model right-sizing, prompt trimming, and enforced p95 budgets — without losing quality.
- Claude vs GPT vs Gemini for Coding in 2026The three frontier model families compared for real coding work — agentic depth, ecosystem fit, context, and cost shape — plus how to actually choose.