AI Model Matrix: Compare AI Models Side by Side

Eleven current models from Anthropic, OpenAI, and Google in one sortable table. Filter by provider, price tier, or context window, and check the calendar below for the three scheduled repricings already on the books. Nothing here calls an API, so nothing can go down or drift while you read it.

A model comparison page built on live API calls breaks the moment a provider renames a model or changes a rate limit. This one doesn’t call anything. Every figure below was read directly from the three providers’ own pricing and model documentation on October 20, 2026, then frozen into the page. The table is plain HTML, so it reads the same with or without JavaScript; filtering and sorting run in your browser and only reorder what is already on the page. It is a dated snapshot, not a live feed: if the verification stamp above is more than a couple of weeks old when you read this, open the source links at the bottom before making a purchasing decision.

DIRECT ANSWER

As of October 20, 2026, every current flagship, mid-tier, and budget model from all three labs sits in the same 1M-token context class — Claude Haiku 5.5 closed Anthropic’s last sub-1M gap on October 7. Context size no longer separates anyone. What separates them is price and pricing behavior: the matrix runs 100x from $0.10 (GPT-6 Luna and Claude Haiku 5.5, short prompts) to $10.00 per million input tokens (GPT-6 Astra and Claude Fable 5.1), and three of the eleven rows carry scheduled or conditional price changes — Gemini 4 Argon’s intro rate expires November 30, GPT-5.6 Sol’s promotional rate expires around November 21, and Gemini 3.8 Flash doubles on January 1, 2027. Use the filters to narrow by what matters for your workload, then read the calendar before you budget.

WHAT CHANGED SINCE OUR SEPTEMBER MATRIX
  • OpenAI: GPT-6 Astra (Sep 3), GPT-6 Sol and Luna (Sep 22), and GPT-6.1 Sol (Sep 29) replaced the entire GPT-5.6 generation. Sol halved the GPT-5.6 promo rate; Luna halved twice in three months ($1.00 → $0.20 → $0.10 input).
  • Anthropic: Claude Opus 5.5 ($4/$20, Sep 22), Sonnet 5.5 ($2/$10, late Sep), and Fable 5.1 moved the Claude 5 family to legacy. Haiku 5.5 launched Oct 7 at $0.10/$0.50 for prompts up to 100K tokens — one-tenth of Haiku 4.5’s input price.
  • Google: Gemini 4 Argon reached paid API general availability on Oct 13 at a $2/$10 introductory rate that doubles December 1.

Filter the matrix

11 current models, 3 providers

specs verified · Oct 20, 2026
Provider
Current Claude, GPT-6, and Gemini models compared by context window, standard API price per 1M tokens, cached-input rate, modality, and release date. Verified October 20, 2026.
Model Context Input $/1M Cached in $/1M Output $/1M Modality Released
GPT-6 LunaOpenAI 1.05M $0.10$0.01$0.50 Text and vision input Sep 22, 2026
Claude Haiku 5.5AnthropicShown: prompts up to 100K tokens ($0.10/$0.50). Over 100K, the whole request reprices to $0.50/$2.50. 1M $0.10$0.01$0.50 Text and vision input Oct 7, 2026
Gemini 3.1 Flash-LiteGoogle 1.05M $0.25$0.025$1.50 Text, vision, and audio input Mar 2026
Gemini 3.8 FlashGoogleIntroductory rate. Doubles to $1.50/$7.50 on Jan 1, 2027, per Google’s pricing page. 1M $0.75$0.075$3.75 Text, vision, and audio input Sep 2026
Gemini 4 ArgonGoogleIntroductory rate through Nov 30; doubles to $4/$20 on Dec 1. Input, thinking, and output share one 1M window. 1.05M $2.00$0.10$10.00 Text, vision, audio, and video input Oct 13, 2026
Claude Sonnet 5.5AnthropicCache reads cut 50% to $0.10 on Oct 7, 2026 — roughly 20% off typical agentic workloads. 1M $2.00$0.10$10.00 Text and vision input Sep 2026
GPT-6.1 SolOpenAIKept GPT-6 Sol’s rates; OpenAI reports near-Astra results on selected coding and agentic evaluations. 1.05M $2.00$0.10$10.00 Text and vision input Sep 29, 2026
Gemini 3.1 Pro PreviewGoogleShown: the ≤200K-token tier. Above 200K, the whole request reprices to $4/$18. 1.05M $2.00$0.20$12.00 Text, vision, and audio input Nov 2025
Claude Opus 5.5Anthropic20% under Opus 5’s launch price, with fewer steps and tokens per task per Anthropic. 1M $4.00$0.20$20.00 Text and vision input Sep 22, 2026
Claude Fable 5.1AnthropicMythos-class, built for long-running autonomous work. Fable 5 moved to legacy. 1M $10.00$0.25$50.00 Text and vision input Sep 2026
GPT-6 AstraOpenAIPrompts over 272K input tokens reprice the whole request to $20/$75 (2x input, 1.5x output). 1.05M $10.00$1.00$50.00 Text and vision input Sep 3, 2026
Vision input Audio input Video input Every model accepts text and returns text.
Showing 11 of 11 models.

Still on the rate cards but superseded: GPT-5.6 Sol at a $4/$20 promotional rate (advertised through at least Nov 21, 2026), GPT-5.6 Terra at $2/$12, and GPT-5.6 Luna at $0.20/$1.20. Anthropic’s Claude 5.0 generation (Haiku 4.5 through Fable 5) moved to legacy. Neither group is in the matrix because neither is the current generation.

How to read this table

Context window is the maximum combined input the model can hold in one request, in tokens. Price is the standard (non-batch, non-Fast-mode) API rate per million tokens; input and output are billed separately, and cached input is the discounted rate for repeated prompt prefixes — at these ratios (90–95% off input on most rows) caching is the single biggest lever on a production bill, so it gets its own column. Batch and Flex processing, which typically halve the standard rates on OpenAI and Google, are left to the source pages to keep the table scannable. Modality flags what the model accepts beyond plain text. Click any column header to sort; click again to reverse.

The amber row notes are repricing triggers, not fine print. OpenAI bills the entire request at 2x input / 1.5x output once it passes 272K input tokens; Google reprices Gemini 3.1 Pro above 200K; Anthropic reprices Haiku 5.5 above 100K. On Gemini 4 Argon, thinking tokens bill as output inside a single 1M window shared with input — a long prompt directly shrinks the generation budget.

ON FIRM GROUND

Every Anthropic figure comes from platform.claude.com docs and Anthropic’s own launch posts (Opus 5.5 and Sonnet 5.5, September 2026; Haiku 5.5, October 7, 2026). Every OpenAI figure — including context windows, which OpenAI now publishes on its model pages — comes from openai.com’s pricing page and developers.openai.com model reference. Every Google figure comes from ai.google.dev/gemini-api/docs/pricing and Google’s October 13 release notes for Argon.

SCOPE, STATED PLAINLY

This matrix covers the current API lineup of Anthropic, OpenAI, and Google — the three labs most teams shortlist first. It excludes open-weights models (DeepSeek V4 Pro, Qwen 3.8-Max, Llama-class) whose self-host and third-party pricing don’t fit a single rate-card column, and preview models without published rates. GPT-6.1 Astra is unreleased: per WSJ and NYT reporting OpenAI has no published model card, endpoint, or price, so no row exists for it.

Price-change calendar: what’s already scheduled

A listed price is now a promise with an expiry date. These are the repricings documented on the providers’ own pages as of October 20, 2026 — budget against the second column, not the first.

DateChangeImpact
Nov 21, 2026 GPT-5.6 Sol promotional rate ($4/$20, cached $0.40) advertised “at least through” this date; the standard rate it returns to is $5/$30. +25% input, +50% output
Nov 30, 2026 Gemini 4 Argon introductory Standard rates end ($2/$10, cached $0.10). Doubles Dec 1, per Google’s pricing page. Batch and Flex halve whichever rate is live. 2x on all token rates
Jan 1, 2027 Gemini 3.8 Flash introductory rates ($0.75/$3.75, cached $0.075) step up to standard ($1.50/$7.50, cached $0.15), per Google’s pricing page. 2x on all token rates
No date set Claude Sonnet 5.5 / Opus 5.5 / Fable 5.1 / Haiku 5.5 carry no scheduled change as of this writing — the only current lineup where the listed price is the whole price, with Anthropic’s Oct 7 cache-read cut going the other direction. —

Myth vs. fact

MYTHFACT
“Only Gemini offers a 1-million-token context window.”That stopped being true in stages through 2026 and ended entirely on October 7: Claude Haiku 5.5 was the last holdout, and every one of the eleven models above now sits in the 1M–1.05M class. Context capacity is table stakes; how a provider bills that capacity is the differentiator — three of eleven rows reprice the whole request once a threshold is crossed.
“The listed price is the price you’ll pay.”Rarely, at the frontier. GPT-6 Astra doubles input and adds 50% to output past 272K tokens — applied to the entire request, not just the overage. Gemini 3.1 Pro reprices above 200K; Haiku 5.5 above 100K. And on Gemini 4 Argon, thinking tokens bill as output, so a “cheap” $2 input rate can pair with a very large output bill on reasoning-heavy work.
“Introductory launch prices last.”Two rows in this matrix carry documented expiry dates. Gemini 4 Argon’s $2/$10 doubles on December 1, 2026; Gemini 3.8 Flash’s $0.75/$3.75 doubles on January 1, 2027. A price published in October can be a different price by the time an annual budget is approved. Check the calendar above before modeling costs.

Frequently asked questions

Which model has the largest context window?

On paper, GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Gemini 3.1 Flash-Lite, and Gemini 3.1 Pro advertise 1,050,000 tokens, and Gemini 4 Argon advertises 1,048,576 — versus 1,000,000 for the Claude family. In practice the practical difference between 1,000,000 and 1,050,000 tokens rarely matters. What matters more: OpenAI’s 1.05M window caps input at 922K per request (128K is reserved for output), Argon’s window is shared between input, thinking, and output, and three providers reprice whole requests past a threshold. “Context window” and “usable context at the listed price” are different numbers.

Why doesn’t this table include reasoning benchmarks or quality scores?

Benchmark scores move faster than pricing and context specs, are measured differently by every lab and third-party tracker, and would make this page stale within weeks rather than months. This matrix is scoped to the facts that are both verifiable and comparatively stable: price, context, modality, and release date. For quality, run your own tasks — cost per accepted result on your workload beats any public leaderboard.

Does a bigger context window mean better output quality?

No. Context window is a capacity limit, not a quality measure. It determines how much text a model can hold in one request, not how well it reasons about that text. A model with a smaller window can still outperform a larger-window model on a given task — and a model that burns 100K tokens of thinking to reach the same answer a cheaper model reaches in 10K is worse for your bill even if it scores higher.

How often will this matrix be updated?

Within 14 days of any change that affects a row — a launch, a repricing, or a retirement. Recent history across the three labs (five launches or repricings in the seven weeks before this version) suggests touching it at least monthly. The verification stamp above shows the last full pass; if it’s more than a few weeks old, trust the source links over the table.

Keep reading on AInvasion

  • Who’s Really Paying for the AI Infrastructure Boom — the spending side of the same 1M-context, multi-vendor race this matrix tracks on the pricing side.
  • Why AI Agent Adoption Stats Keep Contradicting Each Other — most of the models in this matrix are sold on agentic capability; here’s why the adoption numbers behind that pitch don’t agree.
  • Your AI Agent’s Skills Are an Unaudited Software Supply Chain — the models above are increasingly deployed with third-party skills and tools attached; this is what to check before you trust one.
  • The Hidden Bill: How Whole-Request Repricing Changes Long-Context Math — a worked-example deep dive on the 272K, 200K, and 100K thresholds flagged above.

How this was verified: All pricing, context, cached-input, and modality figures were read from the providers’ own pages on October 20, 2026: platform.claude.com/docs (model overviews and pricing), openai.com API pricing plus the developers.openai.com model reference for GPT-6 Astra, and ai.google.dev/gemini-api/docs/pricing plus Google’s October 13 release notes for Gemini 4 Argon. Release dates were cross-checked against each lab’s own launch announcements. Unlike our September version, no figure in this matrix relies on third-party trackers: OpenAI now publishes context windows on its official model pages, so the “corroborated” flags are gone.

Nothing here was estimated from training data or general familiarity with these model families. Every row traces to one of the source links below, checked on the date in the stamp. Where a price is introductory, the expiry date is printed in the row note and in the calendar, sourced from the provider’s own pricing page.

Sources

Tom Morgan covers AI infrastructure and cost for AInvasion, focusing on how provider rate cards and model specs translate into production decisions. He verifies specs directly against each provider’s official documentation before publication rather than relying on secondary trackers, and reprices the matrix within 14 days of any change to a listed rate.

Category: AI Models

How these prices were checked

Every price and context window in the matrix was read from the providers’ own pages on 20 October 2026 and frozen into this page. Nothing here is copied from a third-party tracker, and no provider paid for placement. Prices change without notice, so treat the table as a dated snapshot rather than a quote.

The arithmetic behind the cost comparisons is tested separately. On 19 November 2026 we ran our own calculator through six input states and 66 displayed amounts and published the record, including the two presentation issues the check found. That evidence is in Our Tests.

  • AI Cost Calculator — put your own token volumes against the same rate card.
  • Our Tests — the calculator audit, with the six states and every result.
  • How We Test — what “checked” means here, and what it does not.
  • Methodology — sourcing, labelling and corrections.

What the matrix does not tell you

It does not measure output quality, latency, throughput, regional availability, rate limits on your tier, or how a model handles your data. A cheaper model that fails your task is not cheaper. Use the table to shortlist two or three candidates, then run a task-specific test with your own inputs and a stop condition agreed in advance.