LLM API Pricing, November 2026: The Complete Comparison

This is the November 2, 2026 edition of a comparison first published on June 9, 2026, rebuilt from the rate cards on September 23 and updated on October 19. This edition adds GPT-6.1 Terra, Gemini 4 Argon on Google's paid tier, Gemini 3.6 Flash, Anthropic's Fast mode, OpenAI's Fast and Ultrafast prices for Astra and Terra, and DeepSeek's new Search and Fetch fee. It removes three rows that are no longer on their providers' pricing pages: Gemini 3.7 Flash, Gemini 3.5 Flash Lite and DeepSeek V4-Flash. OpenAI, Anthropic, Google and DeepSeek prices were re-read on November 2. Rows for xAI, Z.ai, MiniMax, Alibaba, Moonshot, Meta and Groq are September 23 prices and were not re-read.

Correction, 2 November 2026: the October 19 edition said Anthropic has no long-context surcharge, but its own table showed Claude Haiku 5.5 billing 5x above 100K tokens. It also counted four models that reprice a whole request past a token threshold; the real count in this comparison is ten, including every GPT-6 model at 272K. It called Gemini 3.8 Flash's increase the first price increase announced with a date, though Sonnet 5's cancelled increase had a date too, and it put Gemini 3.8 Flash's per-task premium at 40% where the measured figure is about 45%. All four are fixed below.

What changed between October 19 and November 2.

• GPT-6.1 Terra (October 26) brings back a middle tier: $1.00 per million input tokens, $0.05 cached and $5.00 output, half of GPT-6.1 Sol. GPT-5.6 Terra had no GPT-6 successor for five weeks; now it does. Details in GPT-6 API pricing.
• Every GPT-6 model has the 272K line. OpenAI's model pages for Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6.1 Terra and GPT-6 Luna all say a prompt over 272K input tokens is billed at 2x input and cache rates and 1.5x output for the full request. Earlier editions showed it on Astra only.
• Gemini 4 Argon is on Google's paid tier at $2.00 / $10.00 with cache reads at $0.10 through November 30, then $4.00 / $20.00 and $0.20 from December 1. It is the second dated Google doubling on this card, after Gemini 3.8 Flash.
• Anthropic sells Fast mode for Claude Opus 5.5 at $8.00 / $40.00, twice standard, and for Opus 5 and Opus 4.8 at $10.00 / $50.00 in research preview.
• DeepSeek started billing Search and Fetch at $0.75 per 1,000 successful requests each, from 00:00 Beijing time on November 1. Its current model names are deepseek-flash (V4.1-Flash) and deepseek-v4-pro; the V4-Flash name is retired but still accepted.
• OpenAI's October 23 retirements happened. The models it shut down and what their suggested replacements cost are in The Replacement Is a Raise. What retires next, including Claude Sonnet 4.5 on November 30, is in the Q4 2026 price and retirement calendar.

What changed between September 23 and October 19.

• Claude Haiku 5.5 (October 7) ties GPT-6 Luna for the cheapest frontier-lab model: $0.10 per million input tokens and $0.50 per million output, cache reads at $0.01. It has a cliff: a prompt over 100K tokens bills every token at 5x, $0.50 / $2.50. Haiku 4.5 at $1.00 / $5.00 is now ten times its successor's short-prompt price. Details in The Ten-Cent Haiku.
• Claude Sonnet 5.5 (September 28) kept Sonnet 5's $2.00 / $10.00, and on October 7 its cache read was cut from $0.20 to $0.10 per million.
• GPT-6.1 Sol (September 29) also lists at $2.00 / $10.00, with cached input at $0.10, half GPT-6 Sol's $0.20. Three model families now share the $2 / $10 sticker, and on the first independent run they cost 10x apart per task. See Three at $2/$10.
• OpenAI added an Ultrafast tier on October 8: GPT-6.1 Sol at $12.00 / $60.00, 6x standard, for roughly 3.8x the output speed. See The Price of a Second.
• OpenAI scheduled a long list of legacy models for retirement on October 23, and several suggested replacements cost 2.5x to 6x more than the model they replace. See The Replacement Is a Raise.
• Gemini 4 Argon was announced September 30 at an introductory $2.00 / $10.00, limited at the time to a single access program. It is now on the paid tier and in the tables below.

The spread got wider, and both ends moved.

In June the most expensive list price in this comparison was $30 per million output tokens (GPT-5.5), and the cheapest frontier-lab model charged $0.40 (Gemini 2.5 Flash-Lite, GPT-4.1 nano). In September the top is $50 (GPT-6 Astra, Claude Fable 5.1) and the bottom is $0.50 (GPT-6 Luna). That is a 100x spread on output, and the same 100x on input, $10.00 against $0.10. Both ends of that range sit inside OpenAI's own lineup, which is the clearest sign yet that vendors now price by task tier, not by brand.

Layer in caching and the gap opens further. A cached GPT-6 Luna input token costs $0.01 per million. An uncached GPT-6 Astra input token costs $10.00. That is 1,000x, from one vendor, on one API. Add DeepSeek V4.1-Flash's off-peak cache hit at $0.003 and it is 3,333x.

Three things changed the shape of the market since June:

1. The flagship rung doubled. GPT-5.5 ($5 / $30) and Claude Opus 4.8 ($5 / $25) were replaced at the top by GPT-6 Astra and Claude Fable 5.1, both at $10 / $50.
2. The mid tier got cheaper, and it stuck. Claude Sonnet 5 launched at an introductory $2 / $10 with a return to $3 / $15 scheduled for September 1; the increase was cancelled. GPT-6 Sol landed at the same $2 / $10 on September 22.
3. The floor is no longer flat. DeepSeek replaced its flat rate card with peak and off-peak windows on August 16, and the new "off-peak" rate is 52% to 507% above the old flat rate. Google shipped Gemini 3.8 Flash at $0.75 / $3.75 with a doubling already scheduled for January 1, 2027.

The question is the same as in June. Not which model is cheapest, but whether your architecture can exploit the spread without paying twice for the wrong answer.

OpenAI: four GPT-6 rungs, a 272K line on all of them, and a promo with a floor date.

GPT-6 arrived in steps. Astra shipped on September 3 as the flagship. Sol and Luna followed on September 22 at half the promotional price of the GPT-5.6 tiers they replace, GPT-6.1 Sol on September 29, and GPT-6.1 Terra on October 26.

Source: OpenAI Developer Community. Source: The Next Web, GPT-6.1 Sol. Source: OpenAI Developer Community, Ultrafast. Source: OpenAI, GPT-6 Astra model page. Source: OpenAI, GPT-6.1 Terra model page. Source: OpenAI, API pricing

Batch and Flex run at 50% of standard on every GPT-6 and GPT-6.1 model. Fast mode, the tier OpenAI renamed from Priority Processing on July 30, is 2x standard. For GPT-6 Sol that gives a flat 4x ladder for the same weights: $1.00 / $5.00 on Flex or Batch, $2.00 / $10.00 standard, $4.00 / $20.00 on Fast.

Two details matter more than the sticker prices. GPT-6.1 Terra replaced GPT-5.6 Terra at half the input price. For five weeks GPT-6 had no Terra, and GPT-6 Sol, which matches GPT-5.6 Terra's $2.00 input and $0.20 cached input at $10.00 output instead of $12.00, made the old Terra strictly dominated. GPT-6.1 Terra at $1.00 / $5.00 is now the cheaper replacement, so the migration is a quality test, not a rename. And GPT-6 Astra cannot be run without reasoning. Sol and Luna keep the reasoning_effort "none" setting; Astra dropped it, so every Astra call carries reasoning tokens billed at the $50.00 output rate. A short answer from Astra is rarely a short bill. The per-task numbers are in GPT-6 API pricing: Luna, Terra, Sol and Astra compared.

The June budget rung, GPT-4.1 nano at $0.10 / $0.40, was among the models OpenAI retired on October 23. GPT-6 Luna is the budget rung now: the same $0.10 input, cached input at $0.01, output $0.10 higher at $0.50, and a 1.05M window.

Anthropic: the Sonnet 5 price cut stuck, and cache reads got cheaper at the top.

Source: Anthropic, Pricing. Source: Anthropic, "Claude Haiku 5.5". Source: Unite.AI on Claude Sonnet 5.5. Source: Anthropic, "Introducing Claude Sonnet 5". Source: VentureBeat

Anthropic now sells Fast mode for its Opus models: Claude Opus 5.5 at $8.00 / $40.00, twice standard, and Claude Opus 5 and Opus 4.8 at $10.00 / $50.00 in research preview. Fast mode pricing applies across the full context window, caching multipliers apply on top, and it is not available with the Batch API.

Claude Sonnet 5 launched on June 30 at $2.00 / $10.00 as introductory pricing, with a return to $3.00 / $15.00 scheduled for September 1. On August 10 Anthropic made the introductory price the standard price. That puts Sonnet 5 at the same list price as GPT-6 Sol, and a third cheaper per token than the Sonnet 4.6 it replaces. It is not a third cheaper per task: Anthropic says Claude 4.7 and later models, Sonnet 5 included, use a tokenizer that produces approximately 30% more tokens for the same text, with the exact increase depending on content. Measure tokens per task before you bank the cut. Details in The Sonnet 5 price hike is cancelled.

Caching is where the September card diverges most from June. The 90%-off cache read that was uniform in June is now a per-model number: 10% of input on Sonnet 5, Haiku 5.5, Haiku 4.5 and Opus 5, 5% on Opus 5.5 and Sonnet 5.5 ($0.20 on a $4.00 input rate, $0.10 on $2.00), and 2.5% on Fable 5.1 ($0.25 on $10.00, a 75% cut from Fable 5's $1.00). Cache writes are unchanged at 1.25x input for a five-minute cache and 2x for a one-hour cache. At Fable 5.1 prices a system prompt read a hundred times costs $12.50 once and $0.25 per read after; the same prefix sent uncached costs $10.00 every time. What that does to a coding agent's bill is in Same sticker, different bill.

Three more things were true in June and are not now. Anthropic has no long-context surcharge on any model except Claude Haiku 5.5: it was removed on March 13, 2026 for Opus 4.6, Sonnet 4.6 and everything newer, and Sonnet 5 charges the same $2.00 / $10.00 at 9,000 tokens or 900,000. Haiku 5.5 is the exception, priced by prompt length: a request over 100,000 tokens, counting cache reads and writes, pays 5x on every token. The Priority Tier is closed to new customers; the service_tier parameter still accepts auto and standard_only, but Priority is no longer something a new account can buy. And the Batch API is still 50% off every model, which puts a batched Haiku 4.5 call at $0.50 / $2.50 and a batched, cache-read Haiku input at $0.05 per million.

Google: Gemini 3.8 Flash is cheap until January 1.

Source: Google, Gemini API pricing. Source: Digital Applied, "Gemini 3.8 Flash Costs the Same Until It Doubles in January". Source: eesel AI, "Gemini 3.8 Flash pricing 2026"

Gemini 3.8 Flash shipped at exactly Gemini 3.7 Flash's price, $0.75 / $3.75, with cached input at $0.075. Both standard numbers double on January 1, 2027, to $1.50 / $7.50. Batch, Flex, Priority and the cache rate double with them; Google now lists cached input at $0.15 from January 1. Gemini 3.6 Flash doubles on the same day, and Gemini 4 Argon doubles a month earlier, on December 1. Gemini 3.7 Flash and Gemini 3.5 Flash Lite no longer appear on Google's pricing page. Dated increases are now routine, and they are the reason to keep a per-model price table with an effective-date column rather than a single number per model.

The sticker is also not the whole story today. Gemini output pricing includes thinking tokens, and 3.8 Flash works harder per task than 3.7 Flash did: about 30% more output tokens and more turns on agentic tasks. Digital Applied measured $0.58 per task against $0.40, about 45% more now, and about 2.9x after January. The full arithmetic is in Gemini 3.8 Flash doubles in January.

The June table's Gemini 3.5 Flash at $1.50 / $9.00 sat above the 3.6, 3.7 and 3.8 Flash line at $0.75 / $3.75, and the current Lite model is Gemini 3.1 Flash Lite at $0.25 input. Gemini 3.1 Pro at $2.00 / $12.00 carries the same 200K-token cliff the June post described for 2.5 Pro: at or above 200K tokens the request bills at $4.00 / $18.00.

DeepSeek: the floor rose, and it now runs on a clock.

DeepSeek ended flat pricing on August 16, 2026 at 16:00 UTC. Every model now has a peak rate and an off-peak rate. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday; off-peak is the other seventeen hours plus weekends. The June post's 50% to 75% off-peak discount between 16:30 and 00:30 UTC no longer exists.

DeepSeek's pricing page lists only these two models; the deepseek-v4-flash name is retired but still accepted. Search and Fetch are billed separately at $0.75 per 1,000 successful requests each, from November 1.

Source: DeepSeek, API pricing. Source: Digital Applied, "Off-Peak LLM Pricing Lands". Source: TechBriefly, "DeepSeek launches V4.1-Flash"

Read the August card against June. V4-Flash was $0.14 / $0.28 flat with a $0.0028 cache hit. Its August 16 off-peak rate was $0.22 / $0.66 with a $0.007 cache hit: 57% more on input, 136% more on output and 150% more on a cache hit, before the peak doubling. V4-Pro's cache hit went from $0.003625 to $0.022, a 507% increase. "Off-peak" is the new list price and "peak" is a 2x surcharge, not a discount off the old card.

DeepSeek still has the cheapest cache hit in this comparison. V4.1-Flash's $0.003 is 2% of its own input rate. But the model it undercuts is now GPT-6 Luna at $0.10 / $0.50, not GPT-4.1 nano, and Luna has no clock. What time-of-day pricing does to a router is in Off-peak LLM pricing lands.

xAI, Zhipu, MiniMax, Alibaba, Moonshot, Meta and the open-weight hosts.

Source: OpenRouter, "The Open Weight Models that Matter". Source: South China Morning Post, August 2026. The Alibaba, Moonshot, MiniMax and Groq rows are the list prices in Nadir's provider catalog, which mirrors the LiteLLM model cost table; the Groq rows were checked against the Groq console on September 15.

Two of these have cliffs of their own. Grok 4.6's reprices every token in a request at or above 200,000 tokens, including the first, at 2x, so a 199K-token prompt and a 201K-token prompt differ by about 2x in cost. Only Claude Haiku 5.5's 5x line at 100K is steeper. The Z.ai coding plan bills "peak" rates for four hours each weekday afternoon, 14:00 to 18:00 UTC+8, and half that at every other hour, including whole weekends.

The June post's Llama 4 host table is retired. Groq no longer carries Llama at all; its open-weight lineup is GPT-OSS and Qwen, and GPT-OSS 20B at $0.075 / $0.30 is the cheapest open-weight row here. The cheapest row of all is Qwen Turbo at $0.05 / $0.20. The wider trend is in the averages: the blended market price of inference fell from $2.04 per million tokens on May 31 to $1.16 to $1.18 in the first week of August, and the open-weight median sits near $0.53 against $2.81 to $3.00 for proprietary models. Full detail in The open-weight price war.

The full comparison: 37 price points across 11 providers.

Sorted by standard output price, cheapest first. DeepSeek rows show the off-peak rate; the peak rate is double.

A dash means the rate is not published in the sources above, not that it is zero.

Three multipliers that move the real bill off the rate card.

1. Tokenizers.

The same text is not the same number of tokens on every model. Anthropic says Claude 4.7 and later models, including Sonnet 5 and the 5.5 family, produce approximately 30% more tokens than Sonnet 4.6 and earlier for the same text, and that the exact increase depends on content. A price cut per token can be a smaller cut per task, and a matching price between two models can hide a 30% gap in what a task actually costs. Compare tokens per task, not dollars per token.

2. Reasoning tokens.

Every provider in this comparison bills thinking at the output rate, and the newest models think more. GPT-6 Astra no longer accepts reasoning_effort "none", so it reasons on every call at $50.00 per million. Gemini 3.8 Flash spends roughly 30% more output tokens per task than 3.7 Flash at the same per-token price. MiniMax M3 writes out its reasoning, which is why it came in closer to an eighth of Opus per task in our own runs than the tenth its per-token price suggests. The output column is a floor, not the bill.

3. Cliffs and clocks.

Ten models in this comparison reprice the whole request once a prompt passes a token threshold:

Haiku 5.5's line is the lowest and the steepest. Providers also reprice by the calendar and the clock: DeepSeek doubles for seven hours every weekday, Google doubles Gemini 4 Argon on December 1 and Gemini 3.8 and 3.6 Flash on January 1, 2027, and GPT-5.6 Sol's $4.00 / $20.00 promo is listed as available at least through November 21, against a $5.00 / $30.00 list. Anthropic has no time-of-day pricing, its introductory Sonnet 5 price became permanent instead of expiring, and its only long-context tier is on Haiku 5.5.

What the list prices imply for a mixed workload.

This is arithmetic on the list prices above, not a forecast. Take an assumed mix and price only the output tokens, since output dominates the bill on every model above.

Against a baseline that sends every call to Opus 5.5 at $20.00, the blended output price is $8.50, 57.5% lower at list. Swap the simple tier to GPT-6 Luna or Claude Haiku 5.5, both $0.50, and it is $5.80, 71% lower, as long as the simple calls stay under Haiku 5.5's 100K-token line. Against a Fable 5.1 baseline at $50.00, the same mix is 83% lower. Change the assumed shares and every number moves: at 30 / 40 / 30 the Anthropic-only mix is $11.50, 42.5% below Opus 5.5.

Two caveats, and they are the whole point. First, the shares are the input, not the output: the fraction of your traffic that genuinely needs the top rung is something you measure, and the blended price follows from it. Second, a cheap answer only saves money if it is accepted. A simple call that has to be re-run on the expensive model costs both prices plus the delay, so the quality bar and the escalation rate belong in the arithmetic too; routing without that check is dead reckoning. That is also why Nadir publishes no universal production savings or quality percentage: those numbers come from outcome-labelled traffic against a declared baseline, per customer, or they are not numbers.

What changed since June, in one list.

• The flagship list price doubled: GPT-5.5 at $5.00 / $30.00 and Claude Opus 4.8 at $5.00 / $25.00 gave way to GPT-6 Astra and Claude Fable 5.1 at $10.00 / $50.00.
• The mid tier converged on $2.00 / $10.00: Claude Sonnet 5's introductory price became permanent, and GPT-6 Sol launched at the same numbers.
• The bottom rung is now GPT-6 Luna at $0.10 / $0.50 with cached input at $0.01.
• Cache reads are no longer a flat 90% off at Anthropic: 10% of input on most models, 5% on Opus 5.5, 2.5% on Fable 5.1.
• DeepSeek's flat rate card became peak and off-peak on August 16. The June off-peak discount is gone, and the new off-peak rate is above the old flat rate.
• Gemini 3.8 Flash shipped at 3.7 Flash's $0.75 / $3.75 with a doubling scheduled for January 1, 2027.
• GPT-5.6 Terra got a successor on October 26: GPT-6.1 Terra at $1.00 / $5.00. GPT-5.6 Sol's promo is listed as available at least through November 21.
• Ten models now reprice the whole request past a token threshold, every GPT-6 model at 272K among them, and Claude Haiku 5.5's 5x line at 100K is the steepest.
• Google scheduled dated doublings for Gemini 4 Argon (December 1) and Gemini 3.8 and 3.6 Flash (January 1, 2027).
• Anthropic's Priority Tier closed to new customers, and Anthropic now sells Fast mode on Opus at 2x. OpenAI renamed Priority Processing to Fast mode on July 30 at the same 2x, and added Ultrafast at 6x.
• The blended market price of inference fell from $2.04 per million tokens at the end of May to $1.16 to $1.18 in early August.

Sources: [OpenAI, API pricing](https://developers.openai.com/api/docs/pricing). [OpenAI, GPT-6 Astra model page](https://developers.openai.com/api/docs/models/gpt-6-astra). [OpenAI Developer Community, "Announcing GPT-6 Sol and GPT-6 Luna"](https://community.openai.com/t/announcing-gpt-6-sol-and-gpt-6-luna/1399925) (September 22, 2026). [Simon Willison, September 22, 2026](https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/). [OpenAI, "Fast mode"](https://developers.openai.com/api/docs/guides/fast-mode) and [OpenAI, "Flex processing"](https://developers.openai.com/api/docs/guides/flex-processing). [Anthropic, Pricing](https://platform.claude.com/docs/en/about-claude/pricing). [Anthropic, "Introducing Claude Sonnet 5"](https://www.anthropic.com/news/claude-sonnet-5). [Anthropic, Service tiers](https://platform.claude.com/docs/en/api/service-tiers). [VentureBeat on Fable 5.1 cache pricing](https://venturebeat.com/technology/anthropics-claude-fable-5-1-and-mythos-5-1-arrive-with-a-75-cost-reduction-for-fable-cache-reads). [Google, Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing). [Digital Applied on Gemini 3.8 Flash](https://www.digitalapplied.com/blog/gemini-3-8-flash-costs-the-same-until-it-doubles-in-january). [eesel AI, Gemini 3.8 Flash pricing](https://www.eesel.ai/blog/gemini-3-8-flash-pricing). [DeepSeek, API pricing](https://api-docs.deepseek.com/quick_start/pricing). [Digital Applied on off-peak pricing](https://www.digitalapplied.com/blog/off-peak-llm-pricing-deepseek-glm-windows-2026). [TechBriefly on DeepSeek V4.1-Flash](https://techbriefly.com/2026/09/11/deepseek-v4-1-flash-api-pricing/). [OpenRouter, "The Open Weight Models that Matter"](https://openrouter.ai/blog/insights/the-open-weight-models-that-matter-june-2026/). [South China Morning Post, August 2026](https://www.scmp.com/tech/tech-trends/article/3363549/enterprise-ai-costs-hit-2026-low-driven-price-wars-chinese-open-source-models-research). Alibaba, Moonshot, MiniMax and Groq list prices from Nadir's provider catalog (LiteLLM model cost table; Groq console checked September 15, 2026). [Anthropic, "Claude Haiku 5.5"](https://www.anthropic.com/claude-haiku-5-5), October 7, 2026. [Unite.AI on Claude Sonnet 5.5](https://www.unite.ai/anthropic-releases-claude-sonnet-5-5-at-unchanged-sonnet-5-pricing/) and Anthropic pricing read October 8, 2026 (Sonnet 5.5 cache reads at $0.10). [The Next Web on GPT-6.1 Sol](https://thenextweb.com/news/openai-gpt-6-1-sol-price-astra-devday). [OpenAI Developer Community, Ultrafast](https://community.openai.com/t/ultrafast-is-rolling-out-today-for-gpt-6-1-sol-in-the-api-codex-and-chatgpt-work/1404475), October 8, 2026. [OpenAI, Deprecations](https://developers.openai.com/api/docs/deprecations), read October 17, 2026. [OpenAI, GPT-6.1 Terra model page](https://developers.openai.com/api/docs/models/gpt-6.1-terra) and [BleepingComputer on GPT-6.1 Terra](https://bleepingcomputer.com/news/artificial-intelligence/openai-launches-gpt-61-terra-with-3x-faster-ai-agents-50-percent-lower-pricing), October 26, 2026. OpenAI, Anthropic, Google and DeepSeek prices re-read November 2, 2026; xAI, Z.ai, MiniMax, Alibaba, Moonshot, Meta and Groq prices as of September 23, 2026.