Fast Tokens Compound: 10-Turn Agent Loops in 12 Seconds

You ask a coding agent to fix a failing build. It does not answer once. It reads the error, opens a file, tries a patch, gets a new result, and keeps moving.

That is the part most benchmarks miss. The unit is not one response. It is the loop.

I ran a small test around that loop: ten turns through the same synthetic debugging session, with the same prompt spine and the same tool-result sequence. The score was simple: how long did the full loop take, and did the model stay useful while moving quickly?

10-turn coding-agent loop 12 seconds to 68 seconds Same loop. Lower is better.

MiMoMiMo-V2.5-Pro-UltraSpeed

12.0s

GLMCerebrasGLM-4.7

21.6s

GLMGMIGLM-5.2 FP8

28.8s

GeminiGemini 3.5 Flash

41.4s

ClaudeOpus 4.8

68.3s

That is the story. MiMo-V2.5-Pro-UltraSpeed made the loop feel live. GLM-4.7 on Cerebras was also fast. GLM-5.2 FP8 landed in the middle on our GMI run. Gemini 3.5 Flash has strong public speed numbers, but took 41.4 seconds here. Opus was much slower, which is fine. Opus is the model I would pay for when the answer matters, not for every scratchpad turn.

The gap looks small until the loop gets long. Ten turns is one debugging pass. A real agent workflow can burn 100 turns across planning, edits, tests, review, and follow-up fixes. Push that to 1,000 turns across a larger automation run, and latency stops being a UX detail. It becomes the work budget.

The table I would actually use

The public score is the Artificial Analysis Intelligence Index. Treat it as a sanity check, not a verdict. The agent loop is the part that tells you how the model feels inside a build system.

Use the sorter to change the question: fastest loop, highest token speed, or highest public intelligence score.

Sort the table

Fastest loop Token speed AA score

Model

10-turn loop

Speed signal

Artificial Analysis score

MiMoMiMo-V2.5-Pro-UltraSpeed

12.0s

481 to 530 tok/s

54 AA Index

GLMCerebrasGLM-4.7

21.6s

626 to 663 tok/s

34 AA Index

GLMGMIGLM-5.2 FP8

28.8s

41 to 90 tok/s

51 AA Index

GeminiGemini 3.5 Flash

41.4s

224 to 271 tok/s

55 AA Index

ClaudeOpus 4.8

68.3s

100 to 190 tok/s

61 AA Index

Setup: ten fixed turns from the same synthetic coding-agent debugging trace. Speed signal is our client-side output tokens per second from short and long probes, after first token. Logos show the provider path. Run configs: MiMo-V2.5-Pro-UltraSpeed default; GLM-4.7 on Cerebras, high; GLM-5.2 on GMI, FP8/high; Gemini 3.5 Flash default; Opus 4.8 xhigh.

Read it three ways. Sort by loop time and MiMo-V2.5-Pro-UltraSpeed wins. Sort by Artificial Analysis score and Opus still leads. Sort by speed and GLM-4.7 on Cerebras jumps up, even though its public intelligence score is lower.

That is the point. A good agent stack is not one leaderboard. It is a routing problem.

Gemini is the better Flash comparison

Gemini 3.5 Flash is the cleaner comparison for this post: public API, Flash-class positioning, 1M context, and strong speed-intelligence numbers. Google describes it as frontier-level intelligence at exceptional speed, and Artificial Analysis gives it a 55 Intelligence Index score with more than 280 output tokens per second.¹

In my loop, it still took 41.4 seconds. That is the useful result. High output-token speed helps, but it does not automatically make the whole agent loop fast. Thinking time, first-token latency, token usage, and provider path all show up in the wall clock.

Why MiMo-V2.5-Pro-UltraSpeed won the loop

MiMo-V2.5-Pro-UltraSpeed is not just “a fast model.” It is a serving stack: FP4 quantization, DFlash speculative decoding, and TileRT runtime work.²

That stack changes the feel of the agent. In the ten-turn run, MiMo-V2.5-Pro-UltraSpeed finished in 12.0 seconds with sub-second mean time to first token. In the short and long probes, it landed around 481 to 530 client-side output tokens per second.

GLM-5.2 is the cleaner reasoning story. It did not race like MiMo, but it produced far fewer output tokens than GLM-4.7 while beating Gemini 3.5 Flash on wall time. Artificial Analysis scores GLM-5.2 at 51 and calls it the leading open-weight model on its Intelligence Index.³ GMI makes the provider story simple: OpenAI-compatible API, GLM-5.2 FP8 endpoint, pay per token.⁴

The GMI speed row is worth an asterisk. Artificial Analysis currently lists GMI FP8 much faster, around 211 output tokens per second on its provider page.⁵ I reran our GMI endpoint after the audit, including a non-thinking pass, and still saw roughly 41 to 49 tokens per second in this test. So I am keeping the measured row, but not treating it as the last word on GMI performance.

Pricing matters, but only as a routing check

Here is the clean version: input and output pricing, normalized to dollars per 1M tokens.

For this exact loop, the rough run costs were still tiny in absolute dollars: MiMo-V2.5-Pro-UltraSpeed about $0.014, GLM-5.2 on GMI about $0.023, Gemini 3.5 Flash about $0.04, and Opus about $0.13.

The exact cents matter less than the shape. Opus cost the most and took the longest. That does not make Opus bad. It means Opus belongs at the end of the chain.

The mistake is asking one model to be the whole system.

The better agent is a router: sprint with the fast model, switch to the careful model when the task gets hard, and save the expensive judge for the answer that actually ships.