Inception’s new Mercury 3 costs less per output token than GPT-6 Luna or Claude Haiku 5.5, but only for now. The Redwood City startup launched the model on Monday, October 19, and the Mercury 3 price on its API is $0.10 per million input tokens and $0.35 per million output tokens during a 50% launch discount. The regular rate is $0.20 and $0.70, and Inception hasn’t said when the discount ends.
Mercury 3 is a diffusion language model, a different way of generating text from the GPT and Claude families, and Inception is selling it on speed: 1,068 tokens per second, it says, against roughly 130 for GPT-6 Luna and 240 for Haiku 5.5. That’s the number developers building voice bots and agent loops will care about. OpenRouter’s live stats, which time real traffic, put Mercury 3 much closer to its rivals.
- Price: $0.10 input and $0.35 output per million tokens ($0.01 cached) during Inception’s 50% launch discount; the list price is $0.20 and $0.70.
- Availability: live since October 19 on Inception’s API (new keys get 100 million free tokens), OpenRouter, Baseten and Vercel. Text in, text out, with a 1-million-token context.
- Speed: Inception claims 1,068 tokens per second. OpenRouter’s live median was 211 on October 20 and 298 on October 21.
- The catch: the discount has no announced end date, the benchmarks are Inception’s own runs at medium effort, and the API defaults to low effort.
Mercury 3 price: cheaper than Luna and Haiku 5.5 only while it’s 50% off
Inception’s model docs list a single rate for Mercury 3, with no long-prompt tier. At the launch price it matches GPT-6 Luna and Claude Haiku 5.5 on input and undercuts both by 30% on output. At list price it costs twice as much as either on input and 40% more on output.
| Per 1M tokens (checked Oct 20) | Input | Output | Cached input |
|---|---|---|---|
| Mercury 3, launch price (50% off) | $0.10 | $0.35 | $0.01 |
| Mercury 3, list price | $0.20 | $0.70 | $0.02 |
| GPT-6 Luna, prompts up to 272K | $0.10 | $0.50 | $0.01 |
| Claude Haiku 5.5, prompts up to 100K | $0.10 | $0.50 | $0.01 |
Long prompts change the math. Anthropic’s Haiku 5.5 docs charge $0.50 in and $2.50 out once a prompt passes 100,000 tokens, a cliff we covered in our Haiku 5.5 price breakdown. OpenAI’s Luna page doubles input and adds 50% to output above 272,000 input tokens. Inception’s docs show no long-prompt tier, so on very long jobs even Mercury 3’s list price is competitive. The three companies use different tokenizers, so equal per-token prices don’t guarantee equal bills.
A diffusion model with a 1M-token window and low effort by default
Most chatbots, including the GPT-6 models, write one token at a time, each waiting on the last. Inception’s Mercury models borrow the diffusion approach behind most image generators instead: they produce many tokens in parallel and refine them over several passes. Inception says Mercury 3 is more than 40 times the size of Mercury 1 and runs about as fast. It hasn’t published Mercury 3’s parameter count.
The docs list a 1,000,000-token context window, up from 260,000 on Mercury 2.5, up to 131,072 output tokens, tool calling and structured outputs. The model takes text only, and the API is OpenAI-compatible. It has four reasoning settings: instant, low, medium and high. If you don’t pick one, Inception’s reasoning docs say Mercury 3 runs on low. Every benchmark score and the 1,068-token speed figure in Inception’s launch table are medium-effort results.
Inception’s table: Mercury 3 wins 8 of 12, loses coding to Haiku 5.5
Inception’s launch post compares Mercury 3 with GPT-6 Luna and Claude Haiku 5.5, all at medium effort. The figures below are a subset of that table. Inception doesn’t say who ran the rival scores, and its own price row shows Mercury 3 at the discounted rate.
| Inception’s table (medium effort) | Mercury 3 | GPT-6 Luna | Claude Haiku 5.5 |
|---|---|---|---|
| Output speed (tokens/s) | 1,068 | ~130 | ~240 |
| GPQA Diamond | 89.23 | 86.36 | 84.51 |
| IFBench | 76.84 | 61.82 | 54.65 |
| SWE-Bench Atlas | 51.08 | 33.33 | 32.26 |
| DeepSWE v1 | 46.31 | 37.76 | 51.62 |
| MRCR v2 (long context) | 75.50 | 62.72 | 78.40 |
| AA-LCR v1.1 (long context) | 77.67 | 78.30 | 74.70 |
Across the 12 capability rows, Mercury 3 leads eight. Haiku 5.5 wins the DeepSWE coding benchmark and the MRCR long-context test, and Luna edges ahead on AA-LCR and a knowledge-accuracy test. Inception also folds these scores into a single composite. On its chart, by our reading, Mercury 3 sits less than half a point above Haiku 5.5. Inception first called that composite an “Intelligence Index,” then said on X it isn’t Artificial Analysis’s index of the same name and renamed it “Model Quality.”
1,068 tokens a second, or about 300?
Inception’s launch post says Mercury 3 hits 1,068 tokens per second “on NVIDIA GPUs,” without naming the chip or the test conditions. In a launch-day Hacker News post signed by Inception’s co-founder and chief scientist, the company was a little more specific, citing “widely available NVIDIA Blackwell GPUs.” OpenRouter, where Inception is the only provider, showed a median throughput of 211 tokens per second on Tuesday morning and 298 tokens per second, with about one second of latency, on Wednesday afternoon. On Wednesday’s OpenRouter stats, the fastest provider for Claude Haiku 5.5 showed 168 tokens per second, and the fastest for GPT-6 Luna showed 88, on OpenAI’s Fast tier, which costs twice the standard rate.
Those aren’t lab conditions. OpenRouter measures real customer requests across every reasoning setting, while Inception’s figure is a medium-effort result under conditions it hasn’t published, and the live numbers move from day to day. Even so, the 4-to-8-times speed advantage in Inception’s launch materials shrinks to about 1.8 times Haiku 5.5 and 3.4 times Luna’s fastest tier on Wednesday’s live data. Artificial Analysis, the independent benchmarking firm, still hadn’t listed Mercury 3 when we checked on October 21; its newest Inception entry is Mercury 2.5.
Early customers quoted by Inception like what they see. Shuhao Zhang, cofounder of TinyFish AI, says in the launch post that Mercury 3 completed 50% of 50 live web-navigation tasks in his company’s agent test setup, “significantly more than GPT-6.1-Sol,” though he didn’t give Sol’s score. Stefano Ermon, Inception’s co-founder, said in the company’s press release that it has “already started training our next model.”
Mercury 2 users have until November 15
The launch comes with a deadline for existing customers. Inception’s models page says Mercury 2 and Mercury 2.5 are no longer available to new users, and existing customers with usage before October 19 keep access until November 15, 2026. Enterprise contracts are handled separately.
That leaves two open questions for anyone thinking of switching: when the 50% discount ends, and what an independent test says about the speed. Inception hasn’t answered the first. We’ll update this story and our LLM News feed when Artificial Analysis or another outside lab publishes numbers.
Frequently asked questions
How much does Mercury 3 cost?
During Inception’s launch discount, Mercury 3 costs $0.10 per million input tokens, $0.35 per million output tokens and $0.01 per million cached input tokens. The regular price is $0.20, $0.70 and $0.02. Inception hasn’t said when the discount ends, and new API keys come with 100 million free tokens.
Is Mercury 3 cheaper than GPT-6 Luna and Claude Haiku 5.5?
At the launch price, yes on output and equal on input: Luna and Haiku 5.5 both charge $0.10 in and $0.50 out for normal-length prompts. At list price Mercury 3 costs more. Haiku 5.5 gets five times pricier above 100,000-token prompts and Luna raises its rates above 272,000, while Inception lists one flat rate.
How fast is Mercury 3?
Inception says 1,068 tokens per second on NVIDIA GPUs at medium reasoning effort. OpenRouter’s live stats showed a median of 298 tokens per second on October 21, 2026 (211 a day earlier), against 168 for Claude Haiku 5.5 and 88 for GPT-6 Luna on their fastest providers. No independent lab benchmark has been published yet.
What is a diffusion language model?
It’s a model that drafts many tokens at once and refines them over several passes, the way image generators refine a picture, instead of writing one token after another like GPT or Claude. Inception says that’s why Mercury models run faster. It hasn’t disclosed Mercury 3’s parameter count.
What is Mercury 3’s context window?
One million tokens, with up to 131,072 output tokens, according to Inception’s docs. It accepts text only and supports tool calling and structured outputs. Reasoning effort can be set to instant, low, medium or high, and the default is low.
Sources
- Inception, “Diffusion at Scale: Introducing Mercury 3” (Oct 19, 2026; launch claims, benchmark table and charts, partner quotes, availability): inceptionlabs.ai/blog/introducing-mercury-3
- Inception docs, Models (Mercury 3 launch and list prices, cached price, 1M context, 131,072 max output, text-only, tool calling and structured outputs): docs.inceptionlabs.ai/get-started/models
- Inception docs, Reasoning Efforts (instant/low/medium/high; Mercury 3 defaults to low): docs.inceptionlabs.ai/capabilities/reasoning-efforts
- Inception, Models page (100 million free tokens; Mercury 2 and 2.5 closed to new users, existing access until Nov 15, 2026): inceptionlabs.ai/models
- Inception press release via Business Wire, “Inception Launches Mercury 3, Setting a New Pareto Frontier in Speed and Intelligence” (Oct 19, 2026, 12:48 PM ET; Stefano Ermon quote, backers): businesswire.com
- @_inception_ai on X, launch post (Oct 19, 2026, 1:37 PM ET): x.com/_inception_ai/status/2112236719068369299; clarification on the “Intelligence Index” name (Oct 19, 2026, 4:54 PM ET): x.com/_inception_ai/status/2112286171514732568
- Hacker News, launch thread post by Inception’s co-founder and chief scientist (Oct 19, 2026; “widely available NVIDIA Blackwell GPUs”): news.ycombinator.com/item?id=50155371
- OpenRouter, Mercury 3 model page and provider stats (median throughput and latency, prices, 131,072 max completion): openrouter.ai/inception/mercury-3; comparison pages for Claude Haiku 5.5 and GPT-6 Luna
- Anthropic, Claude Haiku 5.5 model overview (prices up to and over 100,000-token prompts, tokenizer note): platform.claude.com/docs/en/models/haiku-5-5/overview
- OpenAI, GPT-6 Luna model page (prices, 272,000-token long-prompt rule): developers.openai.com/api/docs/models/gpt-6-luna
- Vercel, “Mercury 3 now available on AI Gateway” (Oct 19, 2026): vercel.com/changelog/mercury-3-now-available-on-ai-gateway
- Artificial Analysis, Inception models page (no Mercury 3 entry as of Oct 21, 2026): artificialanalysis.ai/models/creators/inception
Sources first accessed October 20, 2026, between about 8:40 and 9:00 AM ET. Prices, OpenRouter stats and Artificial Analysis rechecked October 21, 2026, between about 2:30 and 2:45 PM ET.
Last updated: October 21, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
