Recursive Flint is free, and nobody sells it by the token. A month of agent output on a MacBook costs about 17 cents of power, but the cheap paid tier is under $11, and DeepSeek V4.1 Flash scores 35.7 points higher on the one benchmark both report.
Recursive released Flint on Wednesday: a 40B mixture-of-experts coder with 2.5B active parameters, Apache 2.0 weights and a 23 GB 4-bit build that fits on a laptop. Nobody sells it by the token yet, so the question isn't which host is cheapest. It's whether your own machine beats the cheap paid tier.
Photo by Jakub Skafiriak on Unsplash
Short version
Running Flint locally costs almost nothing per token. A month of agent output on a MacBook is about 17 cents of electricity. But the paid models it competes with are already cheap: a heavy month on Claude Haiku 5.5 or DeepSeek V4.1 Flash is under $11, so $1,000 of extra hardware takes most of a decade to pay back. Flint makes sense when code can't leave the machine. On price alone it only wins against Sonnet-class bills, and on the one benchmark both sides report it trails DeepSeek by almost 36 points.
What's actually for sale
Nothing, per token. OpenRouter lists recursive/flint-v1.0:free at $0 in and $0 out, served by Recursive itself in BF16 with the full 262,144-token context. The paid ID, recursive/flint-v1.0, exists but had zero providers when we checked on October 23. We found no listing on Together, Fireworks or DeepInfra. Recursive's model card mentions a hosted API with its own terms and no published price.
The free tier has the usual OpenRouter limits: 20 requests a minute, and 50 a day unless you've ever bought $10 of credits, which lifts it to 1,000. A coding agent spends a request per turn, so 50 a day is one or two real tasks. The $10 unlock is the cheapest way to try Flint seriously, and the credits stay spendable on other models. Read the endpoint's data policy before you point a free tier at proprietary code.
The weights are the real product. Sizes from the Hugging Face repos:
| Build | Download |
|---|---|
| BF16 safetensors | 79.59 GB |
| Unsloth Q8_0 | 42.3 GB |
| Unsloth UD-Q4_K_XL | 25.05 GB |
| Official INT4 Q4_1 GGUF | 25.99 GB |
| Official NVFP4 GGUF | 22.86 GB |
| Official MLX INT4 | 22.58 GB |
| Unsloth UD-Q2_K_XL | 13.56 GB |
NVFP4 is the quant Recursive treats as the release build, which it says lands within 0.7 points of 16-bit on average. LM Studio lists a 24 GB RAM minimum. Plan on more for long contexts.
The bill Flint is trying to replace
Say one developer runs a coding agent most of the working day. Over a month that's roughly 300 million input tokens, 90% of them cache hits as the agent re-reads the same repo, plus 6 million output tokens. Here is that month on the cards Flint is pitched against:
- Claude Haiku 5.5, prompts under 100K$8.70
- DeepSeek V4.1 Flash, off-peak$8.91
- DeepSeek V4.1 Flash, spread evenly$10.77
- MiniMax M3.1 Flash$12.15
- Claude Haiku 5.5, prompts over 100K$43.50
- Claude Sonnet 5.5$147.00
30M fresh input, 270M cache reads, 6M output. Haiku 5.5 under 100K: $3.00 + $2.70 + $3.00. Over 100K the whole request moves to $0.50, $0.05 and $2.50. DeepSeek's even spread puts 35 of 168 weekly hours at its doubled peak rate. Cache-write premiums left out.
Now the local side. At Recursive's headline 125 tokens a second, 6 million output tokens is 13.3 hours of generation. If the laptop pulls 70 W at the wall and power costs 18 cents a kWh, that's 0.93 kWh, or about 17 cents. Prefill adds time and a bit more power, but the total stays well under a dollar.
So the meter goes to zero. What matters is how many months of API the hardware money would have bought. Every $1,000 of extra RAM or a bigger GPU covers about 115 months of the short-prompt Haiku bill above, 93 months of DeepSeek spread evenly, 23 months of Haiku on long prompts, and under 7 months of Sonnet 5.5. If you already own a 32 GB or bigger Mac, the hardware is sunk and Flint really is close to free. If you'd be buying the machine for Flint, the cheap APIs win on cost for years.
Read the benchmark chart carefully
Recursive's chart compares Flint with open models of a similar size, not with the paid APIs above. These are its numbers, BF16 build:
| Benchmark | Flint | Qwen 3.6 | Qwen 3.8 | Muse Glimmer |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 54.9 | 44.9 | 73.0 | 51.7 |
| SWE-bench Pro v2 (in-house) | 89.7 | 73.8 | 94.9 | 86.0 |
| SWE-bench Verified | 74.2 | 73.4 | 75.6 | 76.0 |
| LiveCodeBench v6 | 80.3 | 80.4 | 90.3 | 63.5 |
| IFBench | 89.2 | 67.9 | 79.5 | 77.0 |
From the chart on Recursive's Hugging Face card, read October 23, 2026. Qwen 3.6 is in the chart's sparse MoE group (3B active); Qwen 3.8 and Muse Glimmer are dense models of 27-31B. Some competitor scores were measured by Recursive, the rest are developer-reported.
Three things stand out. Against other 3B-active MoE models, Flint is ahead on agentic coding and instruction following, and level on LiveCodeBench. Against the dense Qwen 3.8 it loses on both headline coding tests, by 18 points on Terminal-Bench. And the 89.7 is on what Recursive calls "SWE-bench Pro v2", which it measured itself. It is not the public SWE-bench Pro, so don't set it next to Sonnet 5.5's 81.3 on the public test.
Terminal-Bench 2.1 is the one number both Flint and a cheap paid model report. DeepSeek put the September build of V4.1 Flash at 90.6. Flint's 54.9 is 35.7 points lower, for a model that costs $0.15 in and $0.60 out off-peak. Both are vendor numbers, and Recursive says it counted solutions found through web retrieval as failures, which other labs may not do. The gap is still too wide to explain away.
Artificial Analysis has no Flint page yet, so there's no independent cost-per-task figure. The closest thing is AtomicAgent's hands-on run on an RTX 5090: 12 of 18 runs passed across six tasks, two tasks failed every time, and llama.cpp's current Flint support was 10-30% slower than it should be.
Speed is the hidden cost
Recursive quotes 125+ tokens a second on an M4 Pro MacBook Pro and 150+ on a 64 GB RTX Spark, both at short contexts. Its own footnote measures speed differently, as an average over a 128K session with prefill left out, so treat the headline as a best case. Artificial Analysis measures Claude Haiku 5.5 at about 235 tokens a second and DeepSeek V4.1 Flash at about 200, and neither of them makes your fans spin.
On an agent loop you wait for every turn, so a model at half the speed that needs more retries can cost you an afternoon. Recursive says so too: Flint "may need more guidance" than larger closed models, and it suggests splitting big tasks into small ones. That's more of your time per task.
Who should bother
Teams that can't send code to a third party now have a strong Apache 2.0 option that fits in 24 GB. That is the case for Flint, and it's a good one. People who already own a 32 GB or bigger Mac and spend more than $40 a month on a Haiku-class API get a nearly free second model for routine edits.
Everyone else should spend the $10 to lift the OpenRouter limit, try Flint on a week of real tasks, and compare the result with what DeepSeek V4.1 Flash or Haiku 5.5 cost for the same week. Our calculator will price your own token mix. At under $11 a month for the cheap tier, saving money is a weak reason to switch. Privacy is a better one.
Sources, read October 23
- Recursive: Introducing Flint - release date, speed claims, NVFP4 quality, guidance caveat
- Flint model card on Hugging Face - parameters, licence, context, benchmark chart, eval settings, file sizes
- OpenRouter Flint listing and free-tier limits - $0 price, provider, request caps
- AtomicAgent - hands-on test results and llama.cpp speed caveat
- TokenCost pricing table - Haiku 5.5, DeepSeek V4.1 Flash, MiniMax M3.1 Flash and Sonnet 5.5 rate cards, with sources