Frontier Model API Pricing, September 2026: Claude vs OpenAI vs Gemini vs DeepSeek

Last updated: September 26, 2026

Official Sources#

All prices verified September 26, 2026 against each provider's live pricing page (Google's pricing page was unreachable again on September 26, so Gemini rows still carry the July 26 verification):

What Changed (September 2026)#

The frontier workhorse tier repriced: GPT-6 Sol (September 22), Claude Opus 5.5 (September 22), and Grok 4.7 (September 21). Three flagship-tier launches inside 48 hours reset the table below. OpenAI's GPT-6 family (pricing verified September 26, 2026) ships Astra at $10/$50, Sol - now the coding-and-agentic flagship - at $2/$10, and Luna at $0.10/$0.50, all with a 1,050,000-token window; GPT-6 Sol at $2/$10 is the first frontier-class model priced below the mid tier, and GPT-6 Luna undercuts DeepSeek on every axis (weekly roundup). Anthropic's Opus 5.5 arrives at $4/$20 with cache reads at $0.20 (a 0.05x multiplier versus the standard 0.1x) and fast mode at $8/$40; it performs at Fable 5.1 level on most work at 40 percent lower cost than Opus 5. SpaceXAI's Grok 4.7 opened at $2/$6 with a fast variant at double the output speed for double the price. The GPT-5.6 family does not go away, but its role changes: GPT-5.6-Sol's $4/$20 promo runs at least through November 21, 2026, now undercut by its own successor.

Claude Fable 5.1 and Mythos 5.1 shipped September 1 at $10/$50 with cache hits cut to $0.25. The new generation keeps Fable 5's sticker rates but drops the cache-read multiplier from 0.1x to 0.025x (Anthropic pricing docs, verified September 26, 2026). For agent loops that re-read a stable system prompt every turn, that is a 75 percent reduction on the line item that dominates agentic input spend.

DeepSeek re-priced and renamed its budget tier (September 10). DeepSeek-V4.1-Flash (552B MoE, asymmetric causal encoder-decoder with 8B active input / 16B output parameters) is the new deepseek-flash model; the old deepseek-v4-flash and deepseek-v4-flash-vision-exp names are retired but temporarily route to V4.1-Flash, and V4-Pro requests have routed to V4.1-Flash at V4.1-Flash rates since September 14 while V4.1-Pro is in the works (release news, pricing, both verified September 26, 2026). Off-peak rates fell to $0.15 input / $0.003 cache hit / $0.60 output (peak 2x), from $0.22 / $0.007 / $0.66.

What Changed (August 2026)#

Claude Sonnet 5 pricing became permanent (August 15). Anthropic removed the promotional framing from Sonnet 5: the $2/$10 per MTok rates are now the standard price, and the previously scheduled September 1 increase to $3/$15 will not occur (Anthropic pricing, verified August 15, 2026). This makes Sonnet 5 structurally the cheapest Claude model above Haiku, not just temporarily. Every "promo ends September 1" note in this post from July is now outdated - treat $2/$10 as the going rate. Sonnet 5 remains the strongest price-performance Claude in the lineup, and the Fable 5 orchestrator playbook economics improve accordingly.

DeepSeek peak/off-peak pricing gets a date and numbers (August 16, 16:00 UTC). DeepSeek's pricing page now lists the exact policy: off-peak rates at half of peak, peak hours 01:00-04:00 and 06:00-10:00 UTC (09:00-12:00 and 14:00-18:00 Beijing Time), effective August 16, 2026 at 16:00 UTC. New rates: V4 Flash at $0.22 input (cache miss) / $0.007 (cache hit) / $0.66 output off-peak, doubling to $0.44 / $0.014 / $1.32 during peak; V4 Pro at $0.66 / $0.022 / $1.98 off-peak, doubling to $1.32 / $0.044 / $3.96 during peak (DeepSeek pricing, verified August 15, 2026). The old flat $0.14/$0.28 Flash pricing dies August 16. The V4 Pro model also updated to version DeepSeek-V4-Pro-0813. The headline: even off-peak, DeepSeek output rates roughly double; the cheap-agent math changes materially for output-heavy loops, and every off-peak-scheduling assumption needs the new numbers.

GPT-5.6 Sol price cut (August 27). OpenAI's pricing page now lists GPT-5.6-Sol at $4/$20 per MTok (cached input $0.40, cache writes $5.00), down from the $5/$30 listed here in previous sections. Long-context rates move to $8/$30, and Fast mode to $8/$40. Promotional pricing is in effect at least through November 21, 2026 (OpenAI pricing, verified August 27, 2026). Sol at $20 output now sits below Claude Opus 5's $25 output and ties GPT-5.5 on input - the first time a 5.6-tier model undercuts Opus on both axes.

What Changed (July 2026)#

GPT-5.6 price cuts landed (July 30). OpenAI cut GPT-5.6 Luna 80% to $0.20/$1.20 per MTok and GPT-5.6 Terra 20% to $2/$12, and renamed Priority Processing to Fast mode (2x pricing, up to 2.5x speed) across the family (announcement, pricing page, both verified July 31, 2026). Luna now undercuts Claude Haiku 4.5 on input by 5x, and Terra ties Claude Sonnet 5's promotional input rate. The lower Luna and Terra rates also flow through to Codex and ChatGPT Work usage counting, so subscription quotas stretch further. Full breakdown in our price-cut analysis and budget tier comparison.

DeepSeek announced peak/off-peak pricing (July 31). DeepSeek's pricing page now warns that the API will adopt a peak/off-peak policy: 2x regular prices during peak hours (09:00-12:00 and 14:00-18:00 Beijing Time, UTC+8), effective date to be announced. The current $0.14/$0.28 V4 Flash rates below are pre-policy; budget models should assume off-peak prices are the floor and peak prices double. The V4 Flash model also updated to version DeepSeek-V4-Flash-0731 with 1M context, 384K max output, and Anthropic-format API support (DeepSeek pricing, verified July 31, 2026).

Claude Opus 5 launched (July 24) at $5/$25 per MTok. Anthropic's new flagship matches Opus 4.8 pricing while delivering near-Fable 5 intelligence on most benchmarks. It tops the Artificial Analysis Intelligence Leaderboard at 61, ahead of Fable 5 (60) and GPT-5.6 Sol (59). Available on the Claude API, Bedrock, Google Cloud, and Foundry. Fast mode (2x base pricing) supported on the Claude API only. Anthropic also lists Claude Mythos 5 (limited availability) at $10/$50, same rates as Fable 5. See the full Opus 5 vs Fable 5 comparison. This effectively resets the frontier tier: you now get Fable 5-competitive intelligence at Opus 4.8 prices.

DeepSeek deprecation deadline passed (July 24). The legacy deepseek-chat and deepseek-reasoner model names stopped serving requests on July 24. Any application still using these model identifiers receives errors. Migrate to deepseek-v4-flash ($0.14/$0.28 per MTok) or deepseek-v4-pro ($0.435/$0.87) if not already done. Our DeepSeek V4 migration guide covers the full transition.

GPT-5.6 family goes GA (July 9). OpenAI's new flagship line ships three tiers: Sol ($5/$30, matching GPT-5.5's rates), Terra ($2.50/$15 at launch, cut to $2/$12 on July 30), and Luna ($1/$6 at launch, cut to $0.20/$1.20 on July 30). All three inherit long-context pricing tiers that double rates above the standard window. Sol and GPT-5.5 now sit at parity on sticker price; routing decisions come down to benchmarks and cache behavior.

Claude Sonnet 5 pricing is now standard ($2/$10). Anthropic's August 15 pricing page lists Sonnet 5 at $2 input / $10 output with no end date and explicitly states the previously announced September 1 increase to $3/$15 will not occur. The promotional window is over in the sense that the price is now permanent. This undercuts every Opus tier and most competitors while staying well below GPT-5.6-Terra on output ($10 vs $12).

The tables below reflect September 26, 2026 verified rates (Gemini and the no-longer-listed GPT-5.5 family rows were last touched on their July/August verification dates).

---

Sticker prices for model APIs have never been this spread out. At the top of the September 2026 table, GPT-5.5-pro charges $180 per million output tokens. At the bottom, GPT-6 Luna charges $0.50 per million output tokens. That is a 360x gap between two models you can call with nearly identical OpenAI-style request bodies; the open-weights floor (DeepSeek V4.1-Flash) sits at $0.60 off-peak, a 300x spread. A month ago the bottom of the table was DeepSeek at $0.66 and the gap was 273x - the floor moved down twice.

This post is the raw API rate card: every current frontier and workhorse model from Anthropic, OpenAI, Google, and DeepSeek, with every number pulled from the live first-party pricing page and stamped with a verification date. If you are pricing coding tool subscriptions like Claude Code or Cursor plans instead, that is a different market with different math - see the companion post on AI coding tool pricing.

Four structural observations matter more than any single row: the September price war that repriced the frontier workhorse tier, OpenAI's family tiering strategy now running from $0.10 Luna to $50 Astra, Anthropic's permanent Sonnet 5 rates and dropping cache multipliers, and the Claude tokenizer change that quietly skews per-MTok comparisons.

The Cross-Provider Price Table#

All prices in USD per million tokens (MTok), standard synchronous API. Anthropic, OpenAI, and DeepSeek rows verified September 26, 2026 against each provider's live pricing page (linked in the Official Sources table above); Gemini rows verified July 26, 2026 (page unreachable on July 31, August 15, August 27, and September 26); DeepSeek rows show the current off-peak rates, with the peak rates in the sub-table below.

GPT-6 long-context rates (over 272K tokens): Astra rises to $20/$75, Sol to $4/$15, Luna to $0.20/$0.75. Opus 5.5 fast mode is $8/$40 and its cache multiplier is 0.05x; Fable 5.1 and Mythos 5.1 cache hits use a 0.025x multiplier; every other Claude model keeps the standard 0.1x.

DeepSeek peak/off-peak rates, in effect since August 16, 2026 at 16:00 UTC, with the September 10 V4.1-Flash repricing (off-peak hours are all times outside 01:00-04:00 and 06:00-10:00 UTC; peak = 2x off-peak; verified September 26, 2026):

Source pages, all verified September 26, 2026 (Gemini verified July 26):

• Claude: platform.claude.com pricing
• OpenAI: developers.openai.com API pricing
• Gemini: ai.google.dev Gemini API pricing
• DeepSeek: api-docs.deepseek.com pricing

A few quick reads off the table. DeepSeek V4.1-Flash output off-peak at $0.60 is roughly 83x cheaper than Claude Fable 5.1 output at $50, and GPT-6 Luna at $0.50 is 100x cheaper (verified September 26, 2026 on the pages above); during DeepSeek's peak hours Flash's advantage drops to about 42x. The September relistings reordered the cost-optimized tier entirely: GPT-6 Luna at $0.10/$0.50 is now the cheapest model on the market on both axes, undercutting DeepSeek V4.1-Flash ($0.15/$0.60 off-peak) and the GPT-5.6 family, and Claude Sonnet 5 at $2/$10 is no longer the only sub-$3 flagship-adjacent rate now that GPT-6 Sol matches it exactly on input and output while sitting at the top of the frontier. Claude Opus 5.5 ($4/$20) replaced Opus 5 ($5/$25) as Anthropic's cost play and carries a 0.05x cache read multiplier, so cached agent loops on Opus 5.5 read at $0.20 - the same line item costs $1.00 on Fable 5.1 and $0.25 there. Gemini caching is still the odd one out: it bills a per-token cache rate plus an hourly storage fee ($1.00 per MTok per hour on the Flash models, $4.50 on 3.1 Pro Preview), so a cache you hold but rarely hit can cost more than no cache.

Three Things the Table Does Not Tell You#

The Claude tokenizer makes per-MTok prices understate cost#

Anthropic's pricing page carries a note that is easy to skim past: Opus 4.7 and later models, including Fable 5, Fable 5.1, and Opus 5.5, use a new tokenizer that "may use up to 35% more tokens for the same fixed text" (Anthropic pricing docs, verified September 26, 2026). The models overview puts the typical figure at roughly 30% (verified September 26, 2026).

This matters for cross-provider comparisons. Fable 5 versus Opus 4.8 is apples to apples, since both use the new tokenizer. But if you compare either against Sonnet 4.5-era baselines, or against another provider using your old token counts, the same prompt now consumes up to a third more billable tokens. The honest comparison requires re-counting your actual prompts with Anthropic's token counting endpoint, not multiplying old counts by new rates. For workload-level math, see Fable 5 production cost modeling.

Anthropic dropped the long-context premium, Gemini did not#

Claude Fable 5.1, Fable 5, Opus 5.5, Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6 now include the full 1M token context window at standard pricing. Anthropic's pricing page states it plainly: a 900K-token request bills at the same per-token rate as a 9K-token request, and caching and batch discounts apply at standard rates across the full window (verified September 26, 2026).

Gemini still tiers. Gemini 3.1 Pro Preview doubles input from $2 to $4 and lifts output from $12 to $18 once a prompt crosses 200K tokens, and the cache rate doubles too (Gemini API pricing, verified July 26, 2026). Gemini 2.5 Pro tiers the same way at the same boundary. So the headline "Gemini Pro is cheaper than Sonnet" claim flips for genuinely long-context work: above 200K tokens, Gemini 3.1 Pro input ($4) exceeds Sonnet 4.6 input ($3), which does not tier at all. OpenAI tiers as well, framed as separate long-context rates: GPT-6 Astra rises from $10 / $50 to $20 / $75, GPT-6 Sol from $2 / $10 to $4 / $15, GPT-6 Luna from $0.10 / $0.50 to $0.20 / $0.75, GPT-5.6-Sol from $4 / $20 to $8 / $30, Terra from $2 / $12 to $4 / $18, Luna from $0.20 / $1.20 to $0.40 / $1.80, and GPT-5.5-pro from $30 / $180 to $60 / $270 (OpenAI pricing, verified September 26, 2026). Of the four providers, only Anthropic and DeepSeek charge one flat rate across the full window.

Caching mechanics differ more than the discounts do#

All four providers discount aggressively, but the shapes are different:

• Anthropic: cache reads at 0.1x base input on most models - except 0.05x on Opus 5.5 and 0.025x on Fable 5.1 and Mythos 5.1 - cache writes at 1.25x (5-minute TTL) or 2x (1-hour TTL). Batch API is a flat 50% off input and output (pricing docs, verified September 26, 2026).
• OpenAI: cached input at 0.1x on most models ($0.50 on GPT-5.5's $5 input), cache writes at 1.25x base input for the GPT-5.6 and GPT-6 families ($0.125 on GPT-6 Luna, $2.50 on Terra, $5.00 on Sol, $12.50 on Astra), no cached rate listed for the pro models, batch at 50% off. Fast mode is 2x standard pricing ($8/$40 on GPT-6 Sol, $4/$20 on GPT-5.6-Sol, $4/$24 on Terra, $0.40/$2.40 on Luna) (OpenAI pricing, verified September 26, 2026).
• Google: per-token cache rate plus an hourly storage fee, batch at 50% off (Gemini pricing, verified July 26, 2026).
• DeepSeek: the most aggressive cache discount on the table. V4.1-Flash off-peak cache hits bill at $0.003 versus $0.15 for misses, a 0.02x multiplier (verified September 26, 2026). The peak/off-peak policy (2x rates during 01:00-04:00 and 06:00-10:00 UTC) has been in effect since August 16, 2026 at 16:00 UTC, and the September 10 repricing cut off-peak cache hits from $0.007 to $0.003 (DeepSeek pricing, verified September 26, 2026).

If your workload is a chat or agent loop that resends a large stable prefix every turn, effective cost is dominated by the cache read rate, not the headline input rate. That reshuffles the table substantially in DeepSeek's and Anthropic's favor.

Head-to-Head by Tier#

Matching models by intended tier rather than by name, OpenAI/Anthropic/DeepSeek rows verified September 26, 2026, Gemini verified July 26:

September's launches shuffled every row except the mid-tier. Max capability: Fable 5.1 and Mythos 5.1 (September 1) hold the $10/$50 slot with cache hits at $0.25, and GPT-6 Astra matches them on sticker price. Frontier workhorse: GPT-6 Sol at $2/$10 is now the tier's cheapest option by a wide margin - it undercuts Sonnet 5's input rate while posting flagship-class scores - and Opus 5.5 at $4/$20 landed 40 percent below Opus 5 while matching Fable 5.1 on most work; Grok 4.7 at $2/$6 (September 21, announcement, verified September 26) sits in the same tier at the lowest output rate of any flagship. Fast and cheap: GPT-6 Luna at $0.10/$0.50 took the cost floor from DeepSeek, whose V4.1-Flash repricing ($0.15/$0.60 off-peak) still undercuts every other non-OpenAI option. Grok is not in the table above because SpaceXAI does not publish a full public rate card in the format the other four do; the $2/$6 API rate is from the official announcement. There is a deeper dive in the Fable 5 cost-per-task analysis, the budget tier comparison, and the September price war roundup.

Decision Guide by Persona#

Side projects and prototypes. GPT-6 Luna at $0.10 / $0.50 is now the cheapest input-and-output combo on the market - closed or open - with DeepSeek V4.1-Flash ($0.15 / $0.60 off-peak) and Gemini 3 Flash Preview ($0.50 / $3.00, with a free tier) close behind. At these prices, model choice is a quality decision, not a budget one. The open-weights angle adds another dimension - see notes on DeepSeek's open-weights economics.

Production apps with steady traffic. The mid-tier triangle is now Claude Sonnet 5 ($2 / $10), GPT-5.6-Terra ($2 / $12), and Gemini 3.1 Pro Preview ($2 / $12 under 200K). Terra's July 30 cut pulled it level with the other two on input, and Sonnet 5's $2/$10 is no longer time-limited, so it wins the mid-tier on output. If your prompts regularly exceed 200K tokens, Gemini's tier flip moves it from cheapest to most expensive.

Agentic and long-horizon workloads. Output tokens dominate agent spend, and output is where the providers diverge most. After the September launches, GPT-6 Sol at $2 / $10 is the first flagship-class model that does not punish output-heavy agent loops - its $10 output undercuts Claude Opus 5.5 ($20), Opus 5 ($25), and Fable 5.1 ($50), and Grok 4.7 at $6 output is the only sub-$10 flagship rate. At the cheap end, GPT-6 Luna's $0.50 output versus Haiku's $5 makes OpenAI the cost winner for input-heavy agent loops - see the budget tier comparison for the per-task math. Fable 5.1 at $50 output only pays off when its quality reduces retries and total tokens. Batch APIs (50% off at Anthropic, OpenAI, and Google) are the easiest structural saving for any agent pipeline that is not latency-sensitive.

Teams routing across providers. With four providers publishing OpenAI-compatible or near-compatible APIs and a 273x output price spread, routing cheap-by-default with escalation on failure is increasingly the rational architecture. The LLM router comparison covers the tooling, and the routing strategies guide has the decision frameworks.

Compliance-constrained teams. Residency costs extra everywhere: Anthropic applies a 1.1x multiplier for US-only inference on Opus 4.6, Sonnet 4.6, and later models, and OpenAI charges a 10% uplift for regional data-residency processing on models released on or after March 5, 2026 (both verified August 27, 2026 on the pricing pages above).

When to Skip the Switch#

Cheaper sticker prices do not automatically mean cheaper bills, and there are real reasons to stay put:

• Your cache is tuned. Prompt caches are provider-scoped and model-scoped. Switching providers means cold caches, re-tuned breakpoints, and a transition period at full input rates that can eat months of theoretical savings.
• Your spend is output-dominated and quality-sensitive. A cheaper model that needs two attempts costs more than an expensive one that needs one. Measure completion rates per task, not dollars per token.
• You depend on provider-specific features. Anthropic's 1-hour cache TTL, Gemini's explicit cache storage model, and DeepSeek's dual OpenAI- and Anthropic-format endpoints are not interchangeable. Migration cost is real engineering time.
• DeepSeek's peak/off-peak policy and model rename are now in effect. The 2x peak-hour multiplier (01:00-04:00 and 06:00-10:00 UTC) landed August 16, 2026 at 16:00 UTC; September's V4.1-Flash repricing cut off-peak rates to $0.15 / $0.60 and renamed the model to deepseek-flash, with deepseek-v4-pro requests routing to V4.1-Flash at V4.1-Flash rates since September 14. Any DeepSeek cost model built on the old $0.28 or even $0.66 output numbers needs re-basing (DeepSeek pricing, verified September 26, 2026).

If none of those apply and your evals show parity, the spread in this table is too large to ignore.

FAQ#

What is the cheapest frontier LLM API in September 2026?#

GPT-6 Luna, at $0.10 per million input tokens and $0.50 per million output tokens, verified September 26, 2026 on OpenAI's official pricing page. Cache reads drop input to $0.01. For the first time the cheapest frontier callable API is a closed-provider model - DeepSeek V4.1-Flash (deepseek-flash) at $0.15 / $0.60 off-peak (cache hit $0.003) is the open-weights floor, but no longer the market floor. DeepSeek peak hours (01:00-04:00 and 06:00-10:00 UTC) double V4.1-Flash to $0.30 / $1.20.

Does Claude charge extra for long context in 2026?#

No. Claude Fable 5.1, Fable 5, Opus 5.5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6 include the full 1M token context window at standard per-token rates, with no premium above 200K tokens (verified September 26, 2026). Gemini 3.1 Pro Preview and Gemini 2.5 Pro still charge higher rates for prompts above 200K tokens. OpenAI's GPT-6 and GPT-5.6 families also tier: in the long-context band GPT-6 Sol rises from $2/$10 to $4/$15, GPT-6 Luna from $0.10/$0.50 to $0.20/$0.75, GPT-5.6-Sol from $4/$20 to $8/$30, Terra from $2/$12 to $4/$18, and Luna from $0.20/$1.20 to $0.40/$1.80.

How much more expensive is Claude Fable 5.1 than GPT-6 Sol?#

Fable 5.1 costs $10 input / $50 output per MTok versus GPT-6 Sol at $2 / $10, so 5x on input and 5x on output. Against GPT-6 Astra ($10 / $50) it is a tie on sticker price with a 4x cheaper cache-read line ($0.25 versus $1.00). Against GPT-5.5-pro ($30 / $180), Fable 5.1 is the cheaper max-capability option by 3x on input and 3.6x on output. All figures verified September 26, 2026.

What is GPT-6 and how does it compare to GPT-5.6?#

GPT-6 is OpenAI's newest flagship family, released September 22, 2026 in three tiers: Astra ($10/$50, the max-capability model), Sol ($2/$10, the coding-and-agentic flagship), and Luna ($0.10/$0.50, the high-volume workhorse), all with a 1,050,000-token context window. GPT-6 Sol undercuts GPT-5.6-Sol ($4/$20) on both axes, and Luna at $0.10/$0.50 is the cheapest model OpenAI has shipped to date. GPT-5.6 stays on the table: Sol's $4/$20 promo runs at least through November 21, 2026, and Terra ($2/$12) and Luna ($0.20/$1.20) keep serving mid-tier traffic.

Why do per-MTok prices understate Claude costs?#

Claude models from Opus 4.7 onward, including Fable 5.1, Fable 5, Opus 5.5, and Sonnet 5, use a new tokenizer that can produce up to 35% more tokens for the same text compared with older Claude models, per Anthropic's pricing documentation (verified September 26, 2026). Token counts measured on older models do not transfer, so per-MTok rate comparisons against pre-4.7 baselines undercount real spend.

Do all providers offer batch discounts?#

Anthropic, OpenAI, and Google all offer a 50% batch discount on input and output tokens (Anthropic and OpenAI verified September 26, 2026, Google verified July 26). DeepSeek's pricing page lists no batch tier, but its standard rates sit below the other providers' batch rates in most tiers anyway.

When does Claude Sonnet 5 promotional pricing end?#

It does not - the promotion became the standard price. Anthropic's pricing page (verified September 26, 2026) states the $2/$10 per MTok rates for Sonnet 5 are now standard, and the previously scheduled increase to $3/$15 on September 1, 2026 did not occur. Budgets built on the July assumption that Sonnet 5 would get more expensive can stand.

Continue Reading#

• GPT-6 Sol vs Claude Opus 5.5 vs Grok 4.7: The September 2026 Price War - the week's three frontier launches, head to head
• Claude Opus 5.5 Release Guide: Fable 5.1-Level Coding at $4/$20 - benchmarks and the OpenCode setup
• Budget AI Coding Models Compared 2026 - GPT-6 Luna vs DeepSeek V4.1-Flash vs Haiku 4.5 head-to-head
• AI Model Routing Strategies for Cost-Effective Coding - how to put these rates to work
• DeepSeek V4 Migration Guide - step-by-step migration from legacy model names, including the September rename