The Fastest AI Models for Coding (And Why Speed Compounds)

Roundup

Speed is the most underrated number on a model card. Benchmark scores get the headlines, prices get the spreadsheets, but tokens per second quietly decides what it feels like to work with a model — and, in agent workflows, how much work actually gets done per hour. This roundup ranks the fastest coding models on our board and makes the case for why speed compounds in ways a leaderboard never shows.

All throughput figures below are the numbers we track on our live board at /benchmarks, alongside each model's benchmark scores from their public sources.

The speed board

Model	Tokens/sec	SWE-bench	LiveCodeBench	In / Out per M	Context
Qwen3.7 Max	204	77.3	87.1	$2.50 / $7.50	1M
Gemini 3.5 Flash	167	79.3	87.6	$1.50 / $9	1M
Gemini 3.1 Pro	147	75.6	88.5	$2 / $12	1M
Claude Haiku 4.5	96	66.6	41.2	$1 / $5	200k
MiniMax M3	95	not published	82.2	$0.30 / $1.20	1M
Grok 4.5	91	not published	87.4	$2 / $6	500k
Claude Sonnet 5	89	85.2	82.4	$3 / $15	1M
GPT-5.3 Codex	87	74.8	87.3	$1.75 / $14	400k

Qwen3.7 Max — the outright speed champion

At 204 tokens per second, Qwen3.7 Max is more than three times faster than the Claude flagships (Fable 5 streams at 67, Opus 4.8 at 60) while posting a 77.3 SWE-bench Verified and an 87.1 LiveCodeBench. That combination — genuinely strong scores at absurd throughput — is why it keeps showing up in our agent loops. Its Terminal-Bench score is not yet published, which is the one open question. Full impressions in our Qwen3.7 Max review.

The Gemini pair — fast and complete

Gemini 3.5 Flash at 167 tokens per second is arguably the best speed-quality package on the board: 79.3 SWE, 76.2 Terminal-Bench, 87.6 LiveCodeBench, 1M context, $1.50/$9. It is the rare fast model with a full published slate and no glaring weakness — see our Flash review. Its bigger sibling Gemini 3.1 Pro runs 147 tokens per second and actually beats Flash on LiveCodeBench at 88.5, though its 75.6 SWE trails. Google is the only lab shipping two genuinely fast models with elite scores.

The 87–96 pack

The middle band is where most builders should shop. Haiku 4.5 (96) and MiniMax M3 (95) are the sprint scouts — cheap, fast, and best kept away from hard autonomous work (Haiku's 35.5 Terminal-Bench is a published warning; M3's is unpublished). Grok 4.5 (91) is three days old as we write this, with an 87.4 LiveCodeBench and no SWE or Terminal-Bench numbers yet. Sonnet 5 (89) and GPT-5.3 Codex (87) are the workhorses: fast enough to feel responsive, strong enough to trust with real implementation.

Why speed compounds

Here is the thesis, and it is the reason this roundup exists: in an agent loop, speed does not add — it multiplies. An agent working a task reads its own output. It writes code, runs it, reads the error, writes a fix, runs it again. Every one of those cycles is gated on generation speed. A model that streams 3x faster completes 3x the iterations per hour at equal quality — which means, over a working session, a slightly weaker fast model can simply out-iterate a slightly stronger slow one. Qwen at 204 tokens per second gets roughly three attempts in the time Fable 5 gets one.

The compounding gets stronger in multi-agent setups, where one agent's output is the next agent's input. A slow model in the middle of a pipeline is a traffic jam: everything downstream idles while it types. This is a big part of why we staff fast models in the high-volume seats of an agent team.

When speed beats benchmark points

Speed wins when the loop is tight and the tasks are tractable. Interactive pairing, where you are reading along and course-correcting — a fast model keeps you in flow, a slow one makes you check your phone. Scout work: repo reads, summaries, triage, test scaffolding, where a few benchmark points change nothing but latency changes everything. And any workflow where the model will iterate against feedback — tests, compilers, linters — because iteration count substitutes for raw capability. In those seats, we will take Flash's 79.3 SWE at 167 tokens per second over a slower 85 nearly every time.

When it does not

Speed loses when the task is hard enough that iteration cannot save you. Deep architectural planning, gnarly multi-file refactors, ambiguous debugging — the problems where a wrong first framing poisons everything after it. There, a fast-but-shallow model just produces wrong answers more quickly, and each failed attempt costs your attention even when the tokens are cheap. Fable 5's 67 tokens per second on a plan that is correct beats 204 on a plan you have to throw away. Slow is smooth, smooth is fast — but only at the top of the difficulty curve.

Two caveats on the numbers

First, throughput is a property of the serving stack, not just the weights. The same model can stream at very different rates across providers, regions, and load conditions, and labs tune their serving over time — so treat these figures as representative of what we observe, not as physical constants. This cuts most interestingly for the open-weight models: DeepSeek V4 Pro's 64 tokens per second is what its first-party API delivers today, but open weights mean a faster host can change that number without changing the model.

Second, tokens per second is not time to useful answer. A terse model at 90 tokens per second can finish before a rambling one at 160, and a reasoning-heavy model may spend a long quiet pause before its first token arrives — latency the throughput figure never shows. In interactive use, time-to-first-token often matters more than streaming rate; in batch agent runs, total tokens generated matters as much as the rate they arrive. The honest metric is task completions per hour on your own workload, and the speed column is the best public proxy for it, not a substitute.

Bottom line

Qwen3.7 Max is the fastest serious coding model in the world right now, Gemini 3.5 Flash is the best fast all-rounder, and the right answer for most teams is both kinds: fast models in the loops, a slow heavyweight on the plans. Live numbers, refreshed nightly, at /benchmarks.