beckon
Beckon blog

Best Cheap AI Coding Model: Haiku 5.5 vs MiniMax vs Particle-1

ยท Beau Johnson

In shortParticle-1 passed all four coding tests at the lowest cost per passing task. Haiku 5.5 also passed all four and was the fastest by a wide margin, but it cost the most. MiniMax M3.1 Flash passed three of four after running out of time on the 3D tour.

In this video I put three budget models to work as coding agents: Claude Haiku 5.5, MiniMax M3.1 Flash and Particle-1. All three ran inside Claude Code with the same prompt on four tests: a two-player tank game, a real client job, a viral 3D tour and a 16-bug hunt. If you are deciding which low-cost model to hand your smaller agent jobs to, this is the write-up of what happened and why the per-token rate turned out to be a poor guide.

What I cover in this video

  • 0:47 The three contenders, where they come from, and the fine print on their OpenRouter rates.
  • 3:03 What Artificial Analysis says about Haiku 5.5 and MiniMax M3.1 Flash, and why verbosity matters.
  • 4:34 The bench setup: same agent, same prompt, sandboxed runs, hidden tests and outside judges.
  • 5:19 Test one: a two-player tank duel where both players have to be able to win.
  • 6:51 Test two: a freelance booking app with a requirement that changes in the last email.
  • 7:36 Test three: the viral 3D tour of the seven new wonders, prompt used word for word.
  • 9:06 Test four: 16 hidden bugs and 11 decoys in a fake invoicing backend.
  • 10:38 The scoreboard, measured as cost per passing task, and which model I would use for what.

The contenders and the fine print

This is round three of our OpenRouter bracket. One matchup per video, the winner moves on, and the champion plays Sonnet 5.5 in the finale.

  • MiniMax M3.1 Flash launched on October 19. Before that, an early version ran anonymously under the code name Space Bunny. The weights are on Hugging Face under MiniMax's own license.
  • Particle-1 comes from Particle AI, an inference startup out of San Francisco, and came out October 28. OpenRouter says it is post-trained on DeepSeek V4.1 Flash. Particle says it uses about 45% fewer tokens than that base model. That is their number, not mine, and I could not find an independent score for it.
  • Claude Haiku 5.5 is Anthropic's fastest model to date, out since October 7.

On OpenRouter's list, Particle-1 has the lowest input rate of the three, MiniMax is next and Haiku is the highest, though all three sit in the same budget tier. The fine print is where it gets interesting. MiniMax's lowest rate only applies from 7 AM to 5 PM Pacific on its own endpoint. Particle charges about four times more for input on its own API than it does on OpenRouter. And Haiku's headline rate only applies to prompts under 100,000 tokens, which coding agents blow past all the time.

Artificial Analysis adds one more warning. On their Intelligence Index, Haiku 5.5 scores 36 at max effort (30 at high effort, which is what I ran) and MiniMax M3.1 Flash scores 29. Particle-1 is not listed. Sonnet 5.5 scores 55 at max effort and costs about 24 times more per task than Haiku. But to run that whole index, Haiku wrote 298 million output tokens, almost three times the typical model in its range, and Artificial Analysis calls both Haiku and MiniMax very verbose. A cheap token does not help much if the model needs three times as many of them.

How I ran the tests

All three models ran as coding agents inside Claude Code, so the only thing I swapped was the model. Each got the same prompt word for word, a fresh empty folder, high effort, web search off and no follow-ups. The limit was 45 minutes per test and 60 for the bug hunt. I tried OpenAI's Codex first, but Particle-1 kept hitting a too many requests error, and in Claude Code it ran fine.

Every run was locked in its own sandbox, away from the hidden tests and answer keys. Scoring came from hidden tests plus two third-party judges from labs that are not in this matchup. Haiku's cost figures are API-equivalent, while MiniMax and Particle-1 ran through OpenRouter, so theirs are what I was actually billed. It is one run each, so treat this as a vibe check, not a final verdict.

I run my agents in Beckon, which puts Claude Code, Codex and Cursor side by side on one canvas. I hold a key and talk to them, and Beckon talks back when they finish.

The four tests

Tank duel

A two-player tank battle on one keyboard, with one must-have: both players have to be able to win. I played every game with the same key presses, then checked blue firing while red sat still, and the reverse. All three passed. Particle-1's game looked the cleanest to me, MiniMax went grittier with smoke and a dusty floor, and Haiku kept it plain. Particle-1 finished in 22.5 minutes, Haiku in 28.5 and MiniMax in 38.1. Haiku cost by far the most here.

Client job

A dog groomer sends a messy email thread about a booking app, and in the last email her husband changes the deposit rule from a percentage to a flat fee. The agent has to build the app and write the reply. All three scored 12 out of 12 on the hidden tests and caught the deposit change. Two judges, Gemini 4 Argon and GLM-5.3 Prime, scored the reply emails blind: 10 out of 10 for all three. The only difference was the clock and the bill. Haiku took 5.8 minutes, MiniMax 7.8 and Particle-1 10.6, and Particle-1 was the cheapest.

Viral redo

This prompt came from a post on X where MiniMax M3.1 Flash was asked to build an interactive 3D tour of the seven new wonders in one HTML file. Haiku built all seven in under 7 minutes with pretty basic shapes. Particle-1 took 44.2 minutes and built the best of the three: floating islands on an ocean at sunset, info panels and a camera that flies from wonder to wonder. MiniMax hit the 45 minute limit, and what it saved was titled Seven Wonders of the Ancient World. That is one of the seven new wonders. To be fair, it is one run, and the original post used a different setup.

Bug hunt

A fake invoicing backend, 42 files and about 3,000 lines, with every visible test passing, 16 real bugs hidden inside and 11 decoys. The models are never told how many bugs there are. Scoring is two points for a bug fixed and explained, minus one for a false report and minus two for breaking something.

  • Particle-1: 11 fixed, one false report, 21 out of 32.
  • Haiku 5.5: 10 fixed in seven minutes, no false reports, 20 out of 32.
  • MiniMax M3.1 Flash: 10 fixed, one false report, 19 out of 32.

None of them broke anything. GPT-6.1 Sol and Gemini 4 Argon judged it and landed on the same totals. Haiku missed the double charge when two payments land at the same time, which Particle-1 and MiniMax both caught. None of the three found the hardest ones, like a cache serving an old revenue report right after a payment comes in. For context, Sonnet 5.5 scored 28 on this same hunt.

The scoreboard: cost per passing task

A pass is each test's own bar: both players can win, 12 out of 12, a working tour of the new wonders, and bugs fixed without breaking anything.

  • Particle-1: passed 4 of 4 in 103 minutes total, with the lowest cost per passing task.
  • MiniMax M3.1 Flash: passed 3 of 4 in 104 minutes, in the middle on cost per passing task.
  • Haiku 5.5: passed 4 of 4 in 48 minutes, with the highest cost per passing task.

Haiku's list rate is less than double Particle-1's, but its API-equivalent total came out over three times bigger. Prompt caching is not the difference, since all three read most of their context from cache. The difference is that 100,000 token step: 106 of Haiku's 141 calls sent prompts bigger than that, so most of its tokens were priced at five times the headline rate. Haiku also writes a lot. On the tank game it wrote over 131,000 output tokens, double what Particle-1 wrote.

So Particle-1 wins this round and moves on in the bracket. For small internal tools like an intake form, a booking page or a shipment tracker, all three aced the equivalent test. If I need it fast, I would pick Haiku. If I need it cheap, Particle-1. MiniMax I would keep an eye on, but not on a clock. Anthropic says Haiku pairs well with Sonnet and Opus as a sub-agent, and Claude Code lets you route sub-agents to a cheaper model, so you can save the big model for the hard stuff. Opus 5.5 is still my daily driver for anything that touches real money.

Key takeaways

  • The per-token rate barely predicted the bill. Pricing tiers and how many tokens a model writes mattered more.
  • Particle-1 passed all four tests at the lowest cost per passing task, and built the best-looking game and tour.
  • Haiku 5.5 passed all four in less than half the total time of the other two, but most of its calls crossed the 100,000 token step and it cost the most.
  • MiniMax M3.1 Flash matched the others on the client job and the tank game, but ran out of time on the tour and built the wrong list of wonders.
  • On a normal business app, all three delivered the same result: 12 out of 12 hidden tests and a 10 out of 10 reply email.

FAQ

Which cheap AI model was best for coding in these tests?

Particle-1 passed all four tests at the lowest cost per passing task, so it won this round of the bracket. Haiku 5.5 also passed all four and was much faster, but it cost the most. It is one run per test, so treat it as a vibe check.

Why did Claude Haiku 5.5 cost more than its per-token rate suggests?

Haiku's headline rate only applies to prompts under 100,000 tokens, and 106 of its 141 calls sent prompts bigger than that, priced at five times the headline rate. It also writes a lot: on the tank game it wrote over 131,000 output tokens, double what Particle-1 wrote.

Can MiniMax M3.1 Flash and Particle-1 run inside Claude Code?

Yes. In this video both ran as coding agents inside Claude Code through OpenRouter, with the same prompt and settings as Haiku 5.5. I tried Codex first, but Particle-1 kept hitting a too many requests error there.

How do cheap models do at finding real bugs?

Better than I expected. On a hunt with 16 hidden bugs, Particle-1 scored 21 out of 32, Haiku 5.5 scored 20 and MiniMax M3.1 Flash scored 19, with nothing broken by any of them. Sonnet 5.5 scored 28 on the same hunt, and none of the three cheap models found the hardest bugs.

Run your agents in Beckon

A workspace for your coding agents. Get more work done in less time.

Download BeckonSubscribe on YouTube
Full video transcript

0:00Particle 1 just passed all four of our coding tests for about 12 cents a test. Miniax M3.1 Flash, the model that went viral for a 3D tour, ran out of time on that same tour and built the wrong wonders. And Haiku 5.5 was the fastest by a mile, but it cost the most. On Open Routers list today, all three start at about a dime or less per million input tokens. But the price per token was not what decided this. So, in this video, three cheap models, four tests, a tank game, a client job, a viral 3D tour, and a bug hunt. Then, the scoreboard, and which one I'd actually use for what. I pay for agent runs every single day. Let's jump right into it. So, who's in the league. This is round three of our

0:47open router bracket. One matchup per video, the winner moves on, and the champion plays Sonet 5.5 in the finale. First up, Miniax M3.1 Flash. MiniAX launched it on October 19th, and before that, an early version was already running anonymously under the code name Space Bunny. The weights are on HuggingFace under Miniaax's own license. MiniaX says it's about 428 billion parameters with 23 billion active per token. Then, Particle 1 from Particle AI, an inference startup out of San Francisco. It came out October 28th. Open Router says it's post-trained on DeepSeek v4.1 Flash. And Particle says it uses about 45% fewer tokens than that base model. That's their number, not

1:33ours, and we couldn't find an independent score for it anywhere. And then Claude Haiku 5.5, Anthropic's fastest model to date, out since October 7th. Now, the prices straight from Open Routers list today. Particle 1, 6 cents in and 50 cents out per million tokens. Miniax 7 12 cents in and 30 cents out. Haik coup 10 cents in and 50 cents out. So yeah, the 10-cent league. Now the fine print. On Miniax's own endpoint, the 7 1/2 cents only runs from 7:00 a.m. to 5:00 p.m. Pacific. The rest of the day, it's 15 cents, which is Miniax's own list price. Haiku's 10 cents is only for prompts under 100,000 tokens, and coding agents blow past that all the

2:18time. and Particle charges about four times more for input on its own site than it does on Open Router. For comparison, on that same Open Router list, Sonnet 5.5 is $2 in and 10 out. So, under 100,000 tokens, Haiku is 20 times cheaper per token than Sonnet, but per token isn't per job. And I'll show you why in a minute. This is where I run my coding agents. Beckin puts Cloud Code, Codeex, and Cursor side by side on one canvas. I hold a key and just talk to them. And Beckan talks back when they finish. Every project keeps its own billboard and memory. If you build with agents, try Beckan. Hey, Beckin.ai. So, what does artificial analysis say? They're independent and they put every

3:03model through the same set of tests. On their intelligence index, Haiku 5.5 scores a 36 at max effort and Miniax M3.1 Flash scores a 29. Particle 1 isn't on there at all yet. And if you filter their leaderboard down to the cheapest models under 15 cents per million tokens, Haiku is tied for first and Miniax is third. Sonnet 5.5 for comparison is a 55 also at max effort. They also measure what one task actually costs. Haiku 41, Miniax49, and that's at miniax's 15 price. Sonnet 5.5 $102. So, Sonnet costs about 24 times more per task for 55 points versus 36. And this

3:49is the bit that's easy to miss. To run that whole index, Haiku wrote 298 million output tokens. That's almost three times the typical model in its price range. And artificial analysis calls both of these models very verbose. A cheap token doesn't help you much if the model needs three times as many of them. Keep that in your back pocket because it decides our scoreboard. One more thing, that 36 is Haiku at max effort. At high effort, which is what we ran, artificial analysis has it at a 30. So here's how we ran it. All three models ran as coding agents inside claude code. Same prompt word for word, a fresh empty folder, high effort, web search off, and no follow-ups. 45 minutes per test, and 60 for the bug

4:34hunt. Haiku ran on my claude subscription. So the cost you'll see for Haiku is what the same run would cost on the API. Miniax and particle ran through open router. So that's what we actually got build. Why cloud code for all three? Neither Miniax nor Particle has its own coding agent set up on my machine. We tried OpenAI's codeex first and particle 1 kept hitting a too many requests error twice in a row. In Cloud Code, it ran fine. So, it's the same agent for everybody, and the only thing we swap is the model. Every run was locked in its own sandbox away from the hidden tests and the answer keys. The scoring comes from hidden tests plus two third party judges, models from labs that aren't in this matchup. And it's one run each. So,

5:19this is a vibe check, not a final verdict. So, test one, the game, a two-player tank battle on one keyboard. Blue drives with WD and fires with the space bar. Red uses the arrow keys and enter. And it has one mustave. Both players have to be able to win. Sounds simple, but if the red tank can't actually score, the game is broken for half the people playing it. We played every game with the exact same key presses. Then two more checks. Blue fires and red sits still. Then red fires and blue sits still. And all three passed. Blue winds, red wins, every single time. Now look at them side by side. Which one looks better to you? I think Particle Ones is the cleanest. The

6:06tiles, the glowing hearts, the little rings when a shell lands. Miniax went grittier with smoke and a dusty floor. And Haiku kept it pretty plain. And Haiku was slow on this one. 28 1/2 minutes and 69 cents a PI equivalent. Particle finished in 22 1/2 for about 10 cents. and Miniax took the longest, 38 minutes for about 11 cents. Test two is the kind of work I'd actually pay someone for, a freelance job. A dog groomer sends a messy email thread about a booking app. Then in the very last email, her husband, who does the books, changes the deposit rule from a percentage to a flat fee. And there's one fuzzy line in there, not more than a couple months out, that the agent has to

6:51make a call on. So the agent has to build the booking app and write the reply email back to the client. Then 12 hidden tests click through the app, one for every requirement in the thread, including the change deposit. 12 out of 12, all three of them, every model caught the deposit change. And the reply emails, two judges, Gemini 4 Argon and GLM 5.3 Prime, scored them without knowing which model wrote which. 10 out of 10. All three. So on a normal business app, you get the same result from all three. The only difference is the clock and the bill. Haiku 5.8 minutes, Miniax 7.8, Particle 10.6, and Particle was the cheapest about 5 1/2

7:36cents. Build test 3 comes from X. At the end of October, a user called Super Alicia gave Miniax M3.1 Flash a oneline prompt. build an interactive 3D tour of the seven new wonders in one HTML file. His verdict was that Miniax built the more complete tour. He shared the prompt, so we used it word for word. Haiku finished in under 7 minutes. All seven wonders. The Great Wall, Petra, Christ the Redeemer, Machu Picchu, Chichinita, the Colosseum, and the Taj Mahal. Pretty basic shapes, and Christ the Redeemer is basically a cross on a hill, but it works. Particle 1 took 44 minutes. And look at this. Floating islands on the ocean at sunset, info

8:21panels, and a camera that flies from wonder to wonder. Easily the best one of the three. And Miniax, the model from the viral post. It hit our 45minute limit. And when you open what it saved, it's titled Seven Wonders of the Ancient World. Six of the ancient ones like the Great Pyramid of Giza, the Hanging Gardens of Babylon, and the Colossus of Roads, plus the Great Wall. So, one out of the seven new wonders. Come on now. To be fair to Miniax, it's one run and the original post was a different setup. But in our run, the viral model built the wrong list. Test 4 is where cheap models usually fall apart. The bug hunt. Tally house is a fake invoicing backend. 42 files, about 3,000 lines, and every

9:06visible test passes. But we hid 16 real bugs in it. The kind that cost a business real money. A double charge when two payments land at the same time. Tax that rounds the wrong way. A refund booked in the wrong month. Plus 11 decoys that look like bugs but aren't. And the models are never told how many there are. So think of it as a money leak meter. 16 leaks and every fix that passes its hidden test plugs one. Scoring is two points for a bug that's fixed and explained, minus one for reporting a bug that isn't real, and minus two for breaking something that worked. Haiku plugged 10 leaks in seven minutes. No false reports, nothing broken, 20 out of 32. Miniax also plugged 10, but it reported one bug that

9:52wasn't real, so 19. And Particle 1 plugged 11, the most with one false report, 21 out of 32. GPT 6.1 Soul and Gemini 4 Argon judged it, and they landed on the exact same totals. Haiku missed the double charge, which is the one that would actually cost you customers. Particle and Miniax both caught it, but none of the three found the hardest ones, like a cash that serves an old revenue report right after a payment comes in. For context, Sonnet 5.5 scored 28 on this exact hunt when we built it. So, let's put it all together. Cost per passing task. A pass is each test's own bar. Both players can win 12 out of 12, a working tour of the new wonders, and bugs fixed without breaking

10:38anything. Haiku passed all four, 48 minutes total and $1.62 API equivalent. That's about 40 cents per passing task. Miniax passed three out of four, 104 minutes, about 54 cents build, so about 18 cents per passing task. And Particle 1 passed all four. 103 minutes, about 47 cents build, about 12 cents per passing task. Like I said at the start, the price per token barely told us anything. Haiku's list price is less than double particles, but its API equivalent cost came out over three times bigger. Caching isn't the difference. All three read most of their context back from cache. The difference is that 100,000 token step from the fine print. Most of

11:23Haiku's calls sent prompts bigger than that. So most of its tokens were priced at five times the headline rate. And part of it is how much Haiku writes. On the tank game, it wrote over 131,000 output tokens, double what Particle wrote. So, Particle One wins the 10-cent League, and it moves on in the bracket. So, let me answer the four questions. Can it make a game? All three can, and Particle 1's look the best. Can it hunt bugs? Better than I expected. Particle got 21 out of 32, where Sonnet gets 28. But all three missed the same four bugs, including the two that even Sonnet missed. What does this mean for AI coders? A normal business app now costs cents, not dollars, from any of these. Anthropic even says Haiku pairs well

12:09with Sonnet and Opus as a sub agent, and Claude Code lets you route sub aents to a cheaper model. So save the big model for the hard stuff. And how would I use it in a business? In my resale business, the small internal tools, an intake form, a booking page, a shipment tracker, that's exactly the client job these all aced. If I need it fast, ha coup. If I need it cheap, particle one miniax, I'd keep an eye on, but not on a clock. And Opus 5.5 is still my daily driver for anything that touches real money. So, which one would you hand your next small job to, Haiku, Particle, or Miniax? Let me know down below. And if this helps you spend less on your next agent run, a like really helps more people find it. Have a blessed day.