Claude, GPT-6 and Gemini 4 side by side: 12 models, prices and measurements
Twelve models from three vendors, one question: which one belongs in which tier of a LiteLLM setup? List prices run from $0.10 to $8 per million input tokens, the Artificial Analysis index from 28 to 65, and the cost per task from $0.15 to $26.54. Everything sits side by side here, with vendor figures and independent measurements kept apart.
A 1:1 comparison sets one vendor's model next to the other vendors' models in each tier, by price class and job. It needs two kinds of numbers: price lists that the vendor sets, and measurements that third parties collect with their own tasks. The tier assignment is ours and not a vendor statement. We did not run any benchmark on these models ourselves. All figures were retrieved from the named pages on 22 November 2026.
How we group the twelve models
Top: Claude Fable 5.2, GPT-6.2 Astra, Gemini 4 Argon. Upper: Claude Opus 5.5, Gemini 4 Carbon. Middle: Claude Sonnet 5.5, GPT-6.2 Sol, Gemini 4 Neon. Small: Claude Haiku 5.5, GPT-6 Luna, Gemini 4 Hydrogen. GPT-6.1 Terra is added as a price neighbor. The groups are not clean: Google has nothing at $8 and $40, and Google itself describes Argon as an efficient frontier reasoning model, yet by price it sits with Opus. OpenAI has nothing at $4 and $20. Prices in the tables are US dollars per million tokens from the vendor pages, standard tier, excluding tax.
List prices by tier
Top
| Model | Input / cache / output | Batch | Note |
|---|---|---|---|
| Claude Fable 5.2 | 8 / 0.20 / 40 | 4 / 20 | context 1M, output 128K |
| GPT-6.2 Astra | 8 / 0.40 / 40 | 4 / 20 | above 272K input: 16 / 0.80 / 60 for the whole request |
| Gemini 4 Argon | 4 / 0.20 / 20 from 1 Dec 2026 (until 30 Nov: 2 / 0.10 / 10) | batch 50% | no free tier |
Upper
| Model | Input / cache / output | Batch | Note |
|---|---|---|---|
| Claude Opus 5.5 | 4 / 0.20 / 20 | 2 / 10 | context 1M, output 128K |
| Gemini 4 Carbon | same as Argon: 4 / 0.20 / 20 from 1 Dec 2026 | batch 50% | promotional price until 30 Nov |
On price, Opus 5.5 matches Carbon and Argon from December. OpenAI has no counterpart here; the nearest is Astra at twice the price.
Middle
| Model | Input / cache / output | Batch | Note |
|---|---|---|---|
| Claude Sonnet 5.5 | 2 / 0.10 / 10 | 1 / 5 | context 1M |
| GPT-6.2 Sol | 2 / 0.10 / 10 | 1 / 5 | above 272K: 4 / 0.20 / 15 |
| Gemini 4 Neon | 3 / 0.15 / 15 | batch 50% | 1.5 times the Sonnet price |
Small
| Model | Input / cache / output | Note |
|---|---|---|
| Claude Haiku 5.5 | 0.10 / 0.01 / 0.50, above 100K prompt: 0.50 / 0.05 / 2.50 | batch 0.05 / 0.25 |
| GPT-6 Luna | 0.10 / 0.01 / 0.50 | above 272K: input doubled, output 1.5 times |
| Gemini 4 Hydrogen | 1.50 / 0.075 / 7.50 | free tier, where Google uses content to improve its products |
| GPT-6.1 Terra | 1 / 0.05 / 5 | price neighbor of Hydrogen and Sonnet |
Hydrogen costs 15 times as much as Haiku and Luna. Terra sits at two thirds of Hydrogen and half of Sonnet 5.5. According to Anthropic, Claude models from 4.7 on count about 30 percent more tokens for the same text, so token counts do not compare across vendors.
Measured independently: index, cost, speed
Sources: the Artificial Analysis (AA) Intelligence Index v5.0.1, the Vals Index (as of 19 November 2026) and Arena Text. The AA index combines nine evaluations. AA lists an effort level per model: Max for all Claude and OpenAI models, High for Argon and Carbon, Xhigh for Neon and Hydrogen. The levels are therefore not equal. Arena lists Opus 5.5 at High and Fable 5.2 at Max, and the three Gemini models at High.
| Tier: model (AA effort) | AA index | AA $/task | Output tokens | Tokens/s | Vals % (rank) | Vals $ | Arena (rank) |
|---|---|---|---|---|---|---|---|
| Top: Fable 5.2 (Max) | 65 | 26.54 | 275M | 108 | 71.31 (1) | 17.94 | 1,502 (8) |
| Top: Astra (Max) | 62 | 4.00 | 65.8M | 46 | 67.66 (8) | 10.68 | not in top 15 |
| Top: Argon (High) | 50 | 9.95 | 113M | 75 | 69.39 (6) | 13.37 | 1,520 (2) |
| Upper: Opus 5.5 (Max) | 59 | 10.04 | 218M | 114 | 70.14 (5) | 24.77 | 1,507 (5) |
| Upper: Carbon (High) | 60 | 7.76 | 205M | 59 | 70.45 (4) | 10.51 | 1,527 (1) |
| Middle: Sonnet 5.5 (Max) | 55 | 10.02 | 335M | 151 | 70.85 (3) | 16.18 | not in top 15 |
| Middle: Sol (Max) | 57 | 1.01 | 60.9M | 67 | 66.06 (10) | 3.18 | not in top 15 |
| Middle: Neon (Xhigh) | 61 | 6.44 | 283M | 104 | 70.87 (2) | 7.40 | 1,512 (3) |
| Small: Haiku 5.5 (Max) | 36 | 0.41 | 298M | 257 | 58.73 (22) | 3.17 | not in top 15 |
| Small: Luna (Max) | 28 | 0.15 | 110M | 107 | 52.32 (34) | 0.45 | not in top 15 |
| Small: Hydrogen (Xhigh) | 54 | 3.16 | 304M | 152 | 68.49 (7) | 9.84 | not in top 15 |
| Neighbor: GPT-6.1 Terra (Max) | 51 | 0.96 | 99.4M | 66 | 60.54 (17) | 2.94 | not in top 15 |
AA and Vals use different task sets, so their costs are not comparable: Fable 5.2 costs $26.54 at AA and $17.94 at Vals, Opus 5.5 costs $10.04 at AA and $24.77 at Vals. The ranking also depends on the source. Fable 5.2 has the highest AA index of the twelve (65) and ranks first at Vals (71.31 percent), but only eighth in Arena (1,502), behind Carbon, Argon, Neon and Opus 5.5. Arena measures human preference on text, AA and Vals measure task completion. Cost per task is the core of Cost per task instead of price per token; its figures ($26.54 and $4.00) match this table.
Two things stand out on the AA pages. Fable 5.2 and Astra have the same list price but need 275 million against 65.8 million output tokens. And AA already prices Carbon and Argon at $4 and $20, the price from December.
What matters to LiteLLM operators
We read the file model_prices_and_context_window.json on the LiteLLM main branch on 22 November 2026. All twelve models are in it. Five points differ from the vendor pages or get expensive if you miss them.
1. Carbon and Argon: 1 December. The entries gemini/gemini-4-carbon, gemini/gemini-4-argon and the vertex_ai/ variants carry $2 and $10, cache $0.10. That is the promotional price. From 1 December the price is $4 and $20. Left unchanged, LiteLLM then counts half of the real amount, and every budget per virtual key lasts twice as long as it should. The -cyber-permissive entries are already at $4 and $20, as they get no promotion. The fix, per the LiteLLM docs: input_cost_per_token and output_cost_per_token under model_info in the model_list entry override the file, so 0.000004 and 0.00002.
2. EU surcharges. Bedrock profiles eu.anthropic.* carry plus 10 percent in the file (Sonnet 5.5: 2.20 and 11). For Azure, azure/eu/gpt-6.2-astra is listed at 9.60 and 48, which is plus 20 percent; OpenAI itself names a 10 percent uplift for regional endpoints on models released on or after 5 March 2026. We did not read the Azure price page. On Vertex, every Gemini 4 line outside global carries plus 10 percent (Neon 3.30 instead of 3). Anthropic's own API knows only inference_geo with global or us; us costs 1.1 times the standard rate, and we found no EU value.
3. Batch on Vertex. vertex_ai/claude-opus-5-5 has 2.50 and 12.50 for batch in the file, Anthropic lists 2 and 10. We did not establish which is right for the route through Vertex.
4. Haiku 5.5 above 100,000 tokens. Anthropic doubles the price once the prompt passes 100,000 tokens (0.50 / 2.50), counting the whole prompt including cache reads. The entries claude-haiku-5-5, global.anthropic.claude-haiku-5-5 and vertex_ai/claude-haiku-5-5 carry this threshold. Only individual entries, such as Snowflake, list a flat price. We did not test the behavior of your version.
5. Fable 5.2: price per task. The entry is right at $8 and $40. The risk is in planning: on the AA tasks Fable 5.2 costs 6.6 times as much as Astra at the same list price.
Examples of price file keys:
| Provider | Key | Price from file (input / output) |
|---|---|---|
| Anthropic | claude-sonnet-5-5 | 2 / 10 |
| Bedrock, EU | eu.anthropic.claude-sonnet-5-5 | 2.20 / 11 |
| Vertex, Claude | vertex_ai/claude-opus-5-5 | 4 / 20 |
| OpenAI | gpt-6.2-sol | 2 / 10 |
| Azure, EU | azure/eu/gpt-6.2-sol | 2.40 / 12 |
gemini/gemini-4-neon | 3 / 15 | |
| Vertex, Gemini | vertex_ai/gemini-4-carbon | 2 / 10 (promotion) |
More on configuration in LiteLLM as a local gateway.
Safety, access, lifecycle
By its own report, OpenAI rates both GPT-6.2 models Critical for cybersecurity and High for biology and chemistry. Astra is described as the first externally released model rated High for AI self-improvement. Google offers Cyber Permissive variants of Carbon and Argon on Vertex, without the promotional price. According to Anthropic, Claude Fable 5.2 requires 30 days of data retention and is not available under zero data retention unless Anthropic expressly authorizes it. Whether your data protection agreement covers that is for you to establish. For Hydrogen, Google uses content in the free tier but not in the paid tier.
Lifecycle: Anthropic will not retire Fable 5.2 before 9 November 2027 or Opus 5.5 before 22 September 2027. Sonnet 4.5 ends on 30 November 2026, replaced by Sonnet 5.5. At OpenAI, gpt-5 and o3 shut down on 11 December 2026, and gpt-4o and gpt-4.1 on 14 January 2027, with Luna as the replacement. We found no shutdown date for Gemini 4.
What we did not measure
We ran no benchmark of these twelve models ourselves. Prices come from vendor pages, indices from Artificial Analysis and Vals, and the ranking from the Arena text leaderboard. We did not find independent scores for SWE-bench, OSWorld and HLE, the release date of GPT-6.1 Terra, system card findings on Sonnet 5.5 and Haiku 5.5, or an official EUR price list for Claude subscriptions. We found no vendor launch videos on the vendors' own pages, so the post embeds none.
If you run LiteLLM, set the Carbon and Argon prices in model_info to $4 and $20 before 1 December, then test two or three models per tier on 20 of your own tasks before fixing a tier.
Further reading
- Anthropic, pricing, model overview, data residency, model deprecations
- OpenAI, pricing, deprecations, GPT-6.2 system card
- Google AI Studio, pricing and Vertex AI, pricing
- Artificial Analysis, Fable 5.2, Vals Index, Arena Text
- LiteLLM, price file and custom pricing
- More from us: AI in production, Corporate LLM
Which model is cheapest in each tier when you count cost per task instead of token price?+
According to Artificial Analysis (as of 22 November 2026): in the top tier GPT-6.2 Astra at $4.00 per task against $26.54 for Claude Fable 5.2; in the upper tier Gemini 4 Carbon at $7.76 against $10.04 for Claude Opus 5.5; in the middle tier GPT-6.2 Sol at $1.01 against $10.02 for Sonnet 5.5; in the small tier GPT-6 Luna at $0.15 against $0.41 for Haiku 5.5. The tasks come from the Artificial Analysis index and are not your tasks.
Why does LiteLLM show too little cost for Gemini 4 Carbon and Argon from 1 December 2026?+
The LiteLLM price file lists both models at $2 input and $10 output, the promotional price until 30 November. Google charges $4 and $20 from 1 December. Without updating the file or adding an override under model_info in your own config, LiteLLM counts half of the real amount.
Does LiteLLM apply the Haiku 5.5 surcharge above 100,000 tokens?+
The claude-haiku-5-5 entries in the price file carry the threshold ($0.50 input, $2.50 output above 100,000 tokens), as on Anthropic's pricing page. Only individual provider entries, such as Snowflake, list Haiku at a flat price. We did not test whether your LiteLLM version applies the threshold in operation.
senn-tech