AI Token Price

Best value AI models at every level

The cheapest model that’s good enough for your work: capability scores from Epoch AI, an independent research group, against the prices we check on every provider’s own page. Updated 11 Nov 2026.

12× less
for the same capability: GPT-6.1 Terra scores level with GPT-6 Astra and costs $213 a month on a coding agent, against $2,598.
56× cheaper
to get a model scoring 160+ than in May: $11.25 per 1M tokens then (GPT-5.5), $0.20 now (Claude Haiku 5.5).
29 of 41
current models scored by Epoch AI. The rest show “Not yet scored”, never a guess.

The cheapest model at each level, 11 Nov 2026

For each level, the cheapest current model that scores level with the reference or higher, priced on our coding-agent example (20 tasks a day, 40 steps each, 90% of repeated input cached).

Capability levelCheapest model level with it or betterCoding agent, per monthvs the reference
Top of the indexLike Claude Fable 5.2 (174) Claude Fable 5.2Score 174 (170–179) · $8.00 / $40.00 $1,521 Nothing cheaper scores as high
Frontier workLike Claude Opus 5.5 (169) GPT-6.1 TerraScore 168 (165–172) · $1.00 / $5.00 $2134× less 4× lessClaude Opus 5.5: $853
Everyday workLike Gemini 3.8 Flash (159) GPT-6 LunaScore 157 (155–159) · $0.10 / $0.50 $25.988× less 8× lessGemini 3.8 Flash: $195
Budget levelLike gpt-oss-120b (140) GPT-6 LunaScore 157 (155–159) · $0.10 / $0.50 $25.9821% less 21% lessgpt-oss-120b: $32.88

“Level” means each score falls inside the other’s range on Epoch AI’s Capabilities Index. What the score means.

Find the cheapest model that’s good enough

Your workload Only show

7 current models score level with or higher than Claude Sonnet 5.5 on Epoch AI’s index. Cheapest for a coding agent (20 tasks a day, 40 steps each, 90% of repeated input cached):

#ModelCapabilityPer monthvs your bar
1 GPT-6.1 TerraOpenAI · $1.00 / $5.00 168 165–172 Level $2132× less 2× less
2 Claude Sonnet 5.5Anthropic · $2.00 / $10.00 166 164–169 Your bar $427 —
3 GPT-6.1 SolOpenAI · $2.00 / $10.00 169 166–172 Level $427 Same
4 Gemini 4 ArgonGoogle · $2.00 / $10.00 167 164–170 Level $427 Same
5 Claude Opus 5.5Anthropic · $4.00 / $20.00 169 166–172 Level $853 2× more
6 Claude Fable 5.2Anthropic · $8.00 / $40.00 174 170–179 Higher $1,521 4× more
7 GPT-6 AstraOpenAI · $10.00 / $50.00 169 166–173 Level $2,598 6× more

Not yet scored by Epoch AI Ministral 3 3B ($23.38), Nemotron 3.5 Lightning ($27.37), Ministral 3 8B ($35.07), Mistral Small 4 ($38.00) and 8 more. Shown for price only; they join the list when Epoch publishes a score.

Costs at list prices checked 13 Oct – 11 Nov 2026. The link in your address bar keeps these settings.Change the agent’s steps and prompts · Price your own traffic

Value map

Each dot is a current model: capability up, cost left. Bars show Epoch’s range. The line joins the models that nothing cheaper matches.

130135140145150155160165170175180 $10$30$100$300$1,000$3,000 Coding agent cost per month (log scale) Capability score gpt-oss-20b: capability 135 (132–138), $20 a monthGPT-6 Luna: capability 157 (155–159), $26 a monthgpt-oss-120b: capability 140 (138–142), $33 a monthClaude Haiku 5.5: capability 160 (158–163), $35 a monthGemma 4 26B-A4B: capability 142 (139–144), $46 a monthQwen3.8 27B: capability 153 (151–155), $50 a monthGLM-5.3 Flash: capability 154 (152–156), $51 a monthDeepSeek V4.1 Flash: capability 159 (157–162), $54 a monthQwen3.6 35B-A3B: capability 144 (142–146), $58 a monthGemini 3.5 Flash-Lite: capability 146 (144–147), $84 a monthGemma 4 31B: capability 144 (142–147), $98 a monthMiniMax M3: capability 148 (146–150), $104 a monthMistral Large 4: capability 155 (152–157), $173 a monthGemini 3.8 Flash: capability 159 (157–161), $195 a monthGPT-6.1 Terra: capability 168 (165–172), $213 a monthDeepSeek V4 Pro: capability 156 (154–158), $244 a monthMuse Spark 1.3: capability 158 (156–160), $335 a monthClaude Sonnet 5.5: capability 166 (164–169), $427 a monthGPT-6.1 Sol: capability 169 (166–172), $427 a monthGemini 4 Argon: capability 167 (164–170), $427 a monthQwen3.7 Plus: capability 148 (146–150), $436 a monthGLM-5.3: capability 157 (155–159), $458 a monthKimi K3.1 Pro: capability 162 (159–165), $520 a monthGemini 3.1 Pro: capability 156 (155–158), $533 a monthGrok 4.7: capability 156 (154–158), $772 a monthKimi K3: capability 159 (157–161), $779 a monthClaude Opus 5.5: capability 169 (166–172), $853 a monthClaude Fable 5.2: capability 174 (170–179), $1,521 a monthGPT-6 Astra: capability 169 (166–173), $2,598 a month gpt-oss-20bGPT-6 LunaClaude Haiku 5.5GPT-6.1 TerraClaude Fable 5.2
Coding agent: 20 tasks a day, 40 steps each, 90% of repeated input cached. Chat app: 50,000 conversations a month, about 2,000 tokens in and 400 out each. Source: Epoch Capabilities Index by Epoch AI, CC BY 4.0, read 11 Nov 2026.

The price of intelligence

The cheapest blended list price (3 input tokens to 1 output, per 1M) of any available model scoring at least 150, 160 and 165 on Epoch AI’s index, at the start of each month. It uses today’s scores and each model’s price on that date.

  • 165+ (now GPT-6.1 Terra, $2.00)
  • 160+ (now Claude Haiku 5.5, $0.20)
  • 150+ (now Claude Haiku 5.5, $0.20)
$0.1$0.3$1$3$10$30 MarAprMayJunJulAugSeptOctNovNow

Show this chart on your site: copy the embed code.

Show the values as a table
Date165+160+150+
1 Mar 2026——$1.13Gemini 3 Flash
1 Apr 2026——$1.13Gemini 3 Flash
1 May 2026—$11.25GPT-5.5$1.13Gemini 3 Flash
1 Jun 2026—$10.00Claude Opus 4.8$1.71Kimi K2.6
1 Jul 2026—$10.00Claude Opus 4.8$1.71Kimi K2.6
1 Aug 2026—$4.50GPT-5.6 Terra$0.45GPT-5.6 Luna
1 Sep 2026$20.00Claude Fable 5.1$4.50GPT-5.6 Terra$0.24GLM-5.3 Flash
1 Oct 2026$4.00Claude Sonnet 5.5$4.00Claude Sonnet 5.5$0.20GPT-6 Luna
1 Nov 2026$2.00GPT-6.1 Terra$0.20Claude Haiku 5.5$0.20Claude Haiku 5.5
12 Nov 2026$2.00GPT-6.1 Terra$0.20Claude Haiku 5.5$0.20Claude Haiku 5.5

How to read this page

  • One independent index. Epoch AI combines results from over 50 benchmarks into one score, with a range for its uncertainty.
  • Level means level. Two models count as level only when each score falls inside the other’s range.
  • No guesses. New models show “Not yet scored” until Epoch publishes a score.
  • Cost on real work. Rankings use cost on a workload, so caching and long-prompt surcharges count.

What the capability score means

Cite this

AI Token Price (2026). Best value AI models at every capability level [Data set]. Updated 11 November 2026. https://aitokenprice.com/best-value-ai-models

Free to reuse under CC BY 4.0 with a link to this page.

Show the price of intelligence on your site

The chart above, kept up to date as prices and scores change. Free to use with the link back. Copy this into your page:

<iframe src="https://aitokenprice.com/embed/value" title="The price of AI intelligence" width="100%" height="460" style="border:0;max-width:720px" loading="lazy"></iframe>
<p style="font-size:12px">From <a href="https://aitokenprice.com/best-value-ai-models">AI Token Price</a>, with capability scores from Epoch AI</p>

Preview the chart · Data: JSON and CSV (CC BY 4.0); capability scores from Epoch AI (CC BY 4.0)

Questions

What is the best value AI model?

It depends on the capability you need. As of 11 Nov 2026, the cheapest current model that scores level with Claude Opus 5.5 on Epoch AI’s Capabilities Index is GPT-6.1 Terra, at $1 / $5 per 1M input and output tokens: about $213 a month for our coding-agent example, against $853 on Claude Opus 5.5.

What is the cheapest AI model that is good enough for coding?

Pick the model you would accept as the bar, choose “Coding agent” in the finder above, and the list shows every model that scores level with it or higher, cheapest first. Coding-specific scores aren’t available for enough models yet, so the ranking uses Epoch AI’s general index; test the shortlist on your own code.

Why rank by cost on a workload instead of price per token?

Because caching, long prompts and output length change the order. A model can be cheaper per token and still cost more on an agent that re-reads its context at every step. The value map lets you switch between price per token and cost on each workload.

Why are some models missing from the list?

Only models with an Epoch AI score can be ranked by capability. Unscored models are listed separately with their prices, and join the ranking when Epoch publishes a score. We never estimate a score.