Our verdict: Claude Haiku 5.5 is the cheapest good AI model in 2026. It costs the same as Luna for prompts up to 100,000 tokens and scores higher on independent tests. GPT-6 Luna is the pick for long prompts and OpenAI stacks, and Gemini 3.8 Flash is best used through Google's free API tier.
The frontier race gets the headlines, but the price war at the bottom of the menu is what changes your bill. OpenAI's GPT-6 Luna launched on September 22 at $0.10 per million input tokens and $0.50 per million output tokens. On October 7, Anthropic matched it with Claude Haiku 5.5, a tenth of what Haiku 4.5 cost.
Claude Haiku 5.5
GPT-6 Luna
Gemini 3.8 Flash
Price and score in one picture
The 3 models at a glance
Claude Haiku 5.5- Highest index score here: 43
- Beats Luna on agent and coding tests
- A tenth of Haiku 4.5's price
The catch: five times the price past 100K tokens.
GPT-6 Luna- Holds its price for long documents
- Function calling and structured outputs
- Biggest library ecosystem
The catch: lowest score of the three, 38.
Gemini 3.8 Flash- Free tier in the Gemini API
- Index score 41, between the other two
- Strong at agent and multi-step work
The catch: paid prices double on January 1.
Price per million tokens
Prices checked October 23, 2026, on each company's own pricing page.
| Model | Released | Input / 1M | Output / 1M | AA Intelligence Index | Best at |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | Oct 7, 2026 | $0.10 ($0.50 over 100K) | $0.50 ($2.50 over 100K) | 43 | Agents, coding, office work |
| GPT-6 Luna | Sep 22, 2026 | $0.10 ($0.20 over 272K) | $0.50 ($0.75 over 272K) | 38 | Long prompts, extraction, routing |
| Gemini 3.8 Flash | Sep 2, 2026 | $0.75 ($1.50 from Jan 1) | $3.75 ($7.50 from Jan 1) | 41 | Free tier, Google stacks |
Index scores are each model's top setting on the Artificial Analysis Intelligence Index.
What the monthly bill looks like
Take a support or document pipeline running 2 billion input tokens and 1 billion output tokens a month, with prompts under 100,000 tokens.
| Model | Monthly bill | With batch (50% off) | Per year |
|---|---|---|---|
| Claude Haiku 5.5 | $700 | $350 | $8,400 |
| GPT-6 Luna | $700 | $350 | $8,400 |
| Gemini 3.8 Flash (2026 price) | $5,250 | $2,625 | $63,000 |
| Gemini 3.8 Flash (from Jan 1, 2027) | $10,500 | $5,250 | $126,000 |
| Claude Haiku 4.5 (legacy) | $7,000 | $3,500 | $84,000 |
Moving from Haiku 4.5 to Haiku 5.5 or Luna saves $6,300 a month, or about $75,000 a year.
What each one offers
| Feature | Haiku 5.5 | GPT-6 Luna | Gemini 3.8 Flash |
|---|---|---|---|
| Long prompts at the base price | Up to 100K | Up to 272K | ✓ One price |
| Image input (vision) | ✓ | ✓ | ✓ |
| Free API tier | ✗ | Free in ChatGPT only | ✓ |
| Batch discount | 50% | 50% | 50% |
Gemini's free tier data can be used to improve Google's products.
Claude Haiku 5.5: the new budget leader
Haiku 5.5 is Anthropic's fastest model, built for high-volume work. On Anthropic's own benchmarks it beats Luna on every test the two share: 39.2% vs 16.4% on Terminal-Bench 4.0 and 72.4% vs 48.9% on the OSWorld 2.1 offline subset. The independent index agrees, 43 to 38.
Use it for lightweight agents, coding helpers, document work and customer chat. Past 100,000 tokens it jumps to $0.50 / $2.50, so send long documents elsewhere.
GPT-6 Luna: cheapest for long prompts
Luna is OpenAI's smallest GPT-6 model and holds its $0.10 / $0.50 price up to 272,000 input tokens. It comes with OpenAI's mature tooling: function calling, structured outputs and the largest ecosystem of libraries.
Use it for tagging, pulling fields out of long documents, routing requests to bigger models and first-pass summaries. It scores lowest here, so add a quality check to anything customer-facing.
Gemini 3.8 Flash: free to start, pricey to scale
Google released 3.8 Flash on September 2, with clear gains in software engineering, agent workflows and multi-step reasoning. It scores 41, ahead of Luna but behind Haiku 5.5, at about seven times their price.
Its real advantage is the free tier in the Gemini API, enough for prototypes and small apps. Paid prices double on January 1, 2027, to $1.50 / $7.50.
What about Haiku 4.5 and the frontier models?
Claude Haiku 4.5 is now a legacy model at $1 / $5, ten times Haiku 5.5. There's no reason to stay on it.
When a wrong answer is expensive (hard coding, long agent runs, legal or financial analysis), send it to Claude Opus 5.5 or GPT-6.1 Sol, compared in our frontier ranking. The best setups use a cheap model to sort requests and a strong model for the hard 10%.
Which one should you use?
The verdict
Claude Haiku 5.5 is the cheapest good AI model in 2026. It matches GPT-6 Luna's $0.10 / $0.50 price for prompts under 100,000 tokens and scores higher; pick Luna for long prompts and Gemini 3.8 Flash for its free tier.
FAQ
What is the cheapest good AI model in 2026?
Claude Haiku 5.5 and GPT-6 Luna tie on price at $0.10 per million input tokens and $0.50 per million output tokens. Haiku 5.5, released October 7, 2026, scores higher on independent tests, so it's our pick for prompts under 100,000 tokens. Luna is cheaper for longer prompts.
Is Gemini 3.8 Flash better than GPT-6 Luna?
On independent tests, yes. It scores 41 on the Artificial Analysis Intelligence Index to Luna's 38. But it costs $0.75/$3.75 per million tokens through December 31, 2026 and $1.50/$7.50 from January 1, 2027, while Claude Haiku 5.5 scores 43 at Luna's price. Gemini's real edge is its free API tier.
Is Claude Haiku 5.5 better than Haiku 4.5?
By a wide margin, and it's cheaper. Haiku 5.5 costs $0.10/$0.50 per million tokens for prompts up to 100,000 tokens, a tenth of Haiku 4.5's $1/$5, and Anthropic's published benchmarks show big gains on coding, agent and office tests. Haiku 4.5 is now a legacy model.
How much can I save with Claude Haiku 5.5 or GPT-6 Luna?
At 2 billion input and 1 billion output tokens a month, GPT-6 Luna and Claude Haiku 5.5 each cost about $700, with Haiku prompts kept under 100,000 tokens. Gemini 3.8 Flash costs about $5,250 at its 2026 price and $10,500 from January 2027. Legacy Claude Haiku 4.5 costs $7,000.
When should I pay for a frontier model instead?
When a wrong answer is expensive: hard coding, long multi-step agent tasks, legal or financial analysis and anything customer-facing where quality is the product. Route those to Claude Opus 5.5 or GPT-6.1 Sol and keep budget models for sorting, extraction and simple chat.