Fast AI models compared: Flash, Haiku and GPT
By Imran Sheikh · 2026-11-12 · Models and tools · 4 min read
Key points
- Claude Haiku 5.5 and GPT-6 Luna share the lowest list price.
- GPT-6.1 Terra scores highest of the fast tier on Artificial Analysis.
- Test at low effort first; speed often matters more than a few points.
Most AI features inside an app don’t need the smartest model available. They need an answer in a second or two, at a price that still works when ten thousand people tap the button. Tagging a journal entry, summarising a note, routing a support message, filling a form from a photo. That’s small-model work.
This post compares the fast tiers from Google, Anthropic and OpenAI as of November 2026. We cover price, speed and quality, and where we’d use each one in a real product.
The fast tiers right now
Each vendor names its small models differently. Anthropic has Claude Haiku 5.5, released October 7, 2026, and described as built for classification, extraction and routing. OpenAI has GPT-6 Luna, its “most efficient model for focused, high-volume tasks”, and GPT-6.1 Terra, described as “a fast, efficient model”. Google has Gemini 3.8 Flash, the cheaper Gemini 3.5 Flash-Lite, and the newer Gemini 4 Hydrogen, pitched at speed for everyday coding and agent work.
| Model | Input / output per 1M tokens | Context window | AA Intelligence Index |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 / $0.50 (prompts up to 100K) | 1M | 30 (high), 36 (max) |
| GPT-6 Luna | $0.10 / $0.50 (up to 272K input) | 1.05M | 21 (high), 28 (max) |
| GPT-6.1 Terra | $1 / $5 (up to 272K input) | 1.05M | 42 (high), 51 (max) |
| Gemini 3.8 Flash | $0.75 / $3.75 until Dec 31, then $1.50 / $7.50 | 1M | 31 (high) |
| Gemini 3.5 Flash-Lite | $0.30 / $2.50 | 1M | 12 |
| Gemini 4 Hydrogen | $1.50 / $7.50 | 1M | Not listed |
Read the price column with care. Claude Haiku 5.5 costs $0.10 input and $0.50 output only for prompts up to 100,000 tokens. Above that it moves to $0.50 and $2.50. GPT-6 Luna has a similar step above 272K input tokens. Gemini 3.8 Flash is on an introductory rate that doubles on January 1, 2027. For features with short prompts, which is most in-app features, the low rates are the ones you’ll pay.
Speed: what users actually feel
Speed has two parts. Time to the first token is the pause before anything appears. Output speed is how fast the rest streams in. For a button that returns a short label, the first number is almost the whole experience.
Artificial Analysis reports a time to first answer token for each model and effort setting. Thinking time counts, so effort matters a lot. As of November 2026, GPT-6 Luna with reasoning off shows 1.74 seconds and GPT-6.1 Terra with reasoning off shows 2.20 seconds. Gemini 3.8 Flash at low effort shows 4.10 seconds. Claude Haiku 5.5 at low effort shows 6.99 seconds, and rises steeply at higher effort. These are their test conditions, with their prompts. Your prompts will give different numbers, so treat this as a direction.
Anthropic’s own docs label Haiku 5.5 the fastest model in its lineup, and its effort parameter defaults to medium. For latency-sensitive features we’d test it at low effort first.
Quality at the small end
The Artificial Analysis Intelligence Index blends nine evaluations across knowledge work, coding, science and general knowledge. On that index GPT-6.1 Terra is the strongest of this group by a clear margin. It scores 42 at high effort, against 30 for Haiku 5.5 and 31 for Gemini 3.8 Flash. It also costs ten times more per input token than Haiku 5.5 or Luna.
That index measures hard, multi-step work. Most small-model features are not hard. Picking one of six tags for a note or pulling a date out of a sentence rarely needs a top score. What it needs is consistency, the right output format every time, and a low price. The only way to know is to run your own examples, which we cover in our post on model metrics.
One more quality signal: LMArena’s text leaderboard, dated November 11, 2026, places Gemini 3.8 Flash ninth overall at 1497. That puts a Flash model in the same voting range as several larger models. For chat-style features, where people judge tone and helpfulness, that matters.
Where each fits
- Tagging, routing and extraction at high volume: Claude Haiku 5.5 or GPT-6 Luna. Same list price, both with a cheap cached input rate. Test both on your own labels.
- Fast replies that need more reasoning, like a short coaching answer or a structured plan: GPT-6.1 Terra, or Claude Sonnet 5.5 at $2 / $10 if you want a Claude model.
- Chat-style features where tone matters and media is involved: Gemini 3.8 Flash. Google also runs live audio and image models on the same API.
- Background jobs that can wait minutes: any of these through a batch API. Anthropic and OpenAI both take 50% off batch requests.
- Agentic coding helpers inside a dev tool: Gemini 4 Hydrogen is built for it, and it appears tenth on LMArena’s WebDev leaderboard with a preliminary score.
Things that change the bill
Small models make it easy to forget the hidden costs. A few to watch.
- Tokenizers differ. Anthropic notes that Haiku 5.5 counts the same text as roughly 30% more tokens than Haiku 4.5. Compare cost per task, not just cost per token.
- Thinking tokens bill as output. A small model at high effort can cost more than a mid model at low effort.
- Caching helps most when you send the same long instructions on every call. Haiku 5.5 cache reads cost $0.01 per million tokens for short prompts.
- Intro prices end. Gemini 3.8 Flash doubles in January. Put the later rate in your forecast.
How we choose
For an in-app feature, we start with the cheapest model that could plausibly work, at the lowest effort. We write twenty real inputs with the answers we expect, run them, and check format and accuracy. If it passes, we ship it and log a sample of live outputs. If it fails, we move up one step, in effort first and then in model. Most features stop at the first or second step. The few that need a bigger model usually tell you so on the first run.
Sources
- Anthropic: Claude Haiku 5.5 overview
- Anthropic: Models overview
- OpenAI: API pricing
- OpenAI: Models
- Google AI Studio: Gemini API pricing
- Google AI Studio: Gemini models
- Artificial Analysis: Model leaderboard
- Artificial Analysis: Intelligence Index methodology
- LMArena: Text leaderboard
- LMArena: WebDev leaderboard