Benchmarks / BulletBench

Measured by Spring Prompt

BulletBench

Name: BulletBench results
Creator: Spring Prompt

When every second of thinking time comes off the clock, which models are fast enough to still make good decisions?

Last updated 12 Oct 2026

Results dated: 29 Sep 2026 to 12 Oct 2026
Results: 15 configurations of 14 models
Unit: ladder Elo
Licence: Spring Prompt original
Sample: 8 games per configuration

Ladder Elo: GPT-6.1 Sol

15 results · ladder Elo, higher is better · lines show the 95% range · ≈ cannot be told apart from the leader. Choose a model to highlight it.Clear highlight

1≈ Gemini 3.8 FlashGoogle 1,126
2≈ GPT-6.1 Sol UltrafastOpenAI 927
3≈ GPT-6 AstraOpenAI 839
4 Gemini 3.5 Flash-LiteGoogle 707
5≈ Gemini 3.1 Pro PreviewGoogle 683
6 GPT-6.1 SolOpenAI 645
7 Claude Haiku 5.5Anthropic 557
7 Claude Opus 5.5Anthropic 557
7 Claude Sonnet 5.5Anthropic 557
7 GPT-6 LunaOpenAI 557
11 Mistral Large 4Mistral AI 555
11 Mistral Medium 3.5Mistral AI 555
13 Claude Haiku 4.5Anthropic 491
14 Claude Fable 5.1Anthropic 358
14 Muse WonderMeta 358

Lightning: strength against thinking time

Each configuration's median time per move at 10 seconds plus 1 a move. Right of the line, a model thinks longer than the increment it earns and the clock drains.

Named: the six bestOther models (hover for names)

Full results

BulletBench Blitz 3+2: ladder Elo, ladder Elo, higher is better
#	Model	Blitz 3+2 · 95% range ladder Elo, higher is better	Lightning 10+1 ladder Elo	Bullet 60s ladder Elo	Move time seconds	Lost on time % of games	Cost per game US dollars	Slowest 10% of moves seconds, 90th percentile	Games at this clock
1≈	Gemini 3.8 FlashGoogle	1,126 949–1,347	–	529	4.0 s	12.5%	$0.0795	7.9 s	8
2≈	GPT-6.1 Sol UltrafastOpenAI	927 657–1,152	197	540	2.1 s	0.0%	$1.10	4.2 s	8
3≈	GPT-6 AstraOpenAI	839 595–1,072	–	197	5.6 s	37.5%	$0.28	14.0 s	8
4	Gemini 3.5 Flash-LiteGoogle	707 458–933	892	780	0.8 s	0.0%	$0.0087	–	–
5≈	Gemini 3.1 Pro PreviewGoogle	683 487–1,130	–	–	6.2 s	50.0%	$0.21	16.2 s	8
6	GPT-6.1 SolOpenAI	645 0–890	–	33	4.8 s	37.5%	$0.0767	16.0 s	8
7	Claude Haiku 5.5Anthropic	557 370–727	–	–	6.2 s	37.5%	$0.0164	10.2 s	8
7	Claude Opus 5.5Anthropic	557 370–727	–	–	7.5 s	62.5%	$0.18	13.6 s	8
7	Claude Sonnet 5.5Anthropic	557 370–727	–	–	8.1 s	37.5%	$0.12	14.0 s	8
7	GPT-6 LunaOpenAI	557 370–727	–	–	7.0 s	50.0%	$0.0072	20.2 s	8
11	Mistral Large 4Mistral AI	555 392–727	249	–	0.8 s	0.0%	$0.0082	2.1 s	8
11	Mistral Medium 3.5Mistral AI	555 396–765	529	385	0.5 s	0.0%	$0.0261	0.8 s	8
13	Claude Haiku 4.5Anthropic	491 299–659	33	445	1.3 s	0.0%	$0.0318	1.9 s	8
14	Claude Fable 5.1Anthropic	358 0–619	–	–	5.4 s	37.5%	$0.38	18.2 s	8
14	Muse WonderMeta	358 0–600	–	–	7.1 s	75.0%	$0.27	–	–

Swipe the table sideways for more columns.

Ranks follow the score as shown, so equal numbers share a rank. ≈ marks results whose 95% range overlaps the leader's: they cannot be told apart from it. Each model runs at its provider's default reasoning setting and, where the provider offers a faster one, again at its fastest setting, listed separately. Blitz 3+2 was run at default settings only. Equal ratings are not a glitch: the ladder's ratings come in steps (see Honest caveats below).

Progress over time

Ladder Elo, plotted at each model's release date, higher is better. Each model counts once, at its best setting. Scores are today's, not launch day's. Hover or tap a mark for its model.

Best score so far

Every model at its release date. The line moves each time a new release beats every earlier one, and each new best is named where there is room.

Best score by lab

Each lab's best so far, stepping whenever one of its releases beats its own earlier models.

GoogleOpenAIAnthropicMistral AIMeta

Too slow for the clock

These configurations lost their first four games on time, so they were stopped there and are not rated for that clock. Each shows its median time per move and the time that 1 move in 10 took or exceeded: a median under the increment still loses on time when the slow tail is long. Being too slow is the result, not a fault.

Bullet 60s

Claude Fable 5.1 median 4.2 s, 1 in 10 moves 6.3 s+
Claude Fable 5.1 (low reasoning) median 4.3 s, 1 in 10 moves 9.3 s+
Claude Haiku 5.5 median 4.6 s, 1 in 10 moves 8.1 s+
Claude Haiku 5.5 (low reasoning) median 3.6 s, 1 in 10 moves 9.5 s+
Claude Opus 5.5 median 3.1 s, 1 in 10 moves 5.8 s+
Claude Opus 5.5 (low reasoning) median 3.1 s, 1 in 10 moves 5.3 s+
Claude Sonnet 5.5 median 3.0 s, 1 in 10 moves 9.4 s+
Claude Sonnet 5.5 (low reasoning) median 2.6 s, 1 in 10 moves 5.0 s+
DeepSeek V4 Pro 0423 median 5.0 s, 1 in 10 moves 27.8 s+
DeepSeek V4.1 Flash (low reasoning) median 1.9 s, 1 in 10 moves 10.6 s+
Gemini 3.1 Pro Preview median 4.9 s, 1 in 10 moves 8.4 s+
Kev 4B median 1.9 s, 1 in 10 moves 15.4 s+
Muse Spark 1.3 median 7.0 s, 1 in 10 moves 28.1 s+
Muse Wonder median 3.6 s
Mistral Large 4 median 0.8 s, 1 in 10 moves 2.9 s+
Kimi K3 median 3.2 s, 1 in 10 moves 33.8 s+
GPT-6 Luna median 4.6 s, 1 in 10 moves 10.8 s+
GPT-6 Sol median 4.2 s, 1 in 10 moves 9.4 s+
Qwen3.8 Flash median 3.8 s, 1 in 10 moves 16.9 s+
Qwen3.8 Max (0902) median 4.9 s, 1 in 10 moves 22.8 s+
Qwen3.8 Max (0902) (minimal reasoning) median 4.7 s, 1 in 10 moves 10.3 s+
Grok 4.7 median 4.7 s, 1 in 10 moves 17.7 s+
Grok 4.7 (low reasoning) median 5.4 s, 1 in 10 moves 15.5 s+
MiMo-V2.6-Flash median 4.0 s, 1 in 10 moves 11.7 s+
GLM 4.7 Flash median 9.7 s, 1 in 10 moves 23.1 s+
GLM 5.3 Flash (low reasoning) median 2.2 s, 1 in 10 moves 5.3 s+

Lightning 10+1

Claude Fable 5.1 median 4.4 s, 1 in 10 moves 6.7 s+
Claude Fable 5.1 (low reasoning) median 4.1 s, 1 in 10 moves 6.7 s+
Claude Haiku 5.5 median 1.9 s, 1 in 10 moves 4.8 s+
Claude Haiku 5.5 (low reasoning) median 1.9 s, 1 in 10 moves 2.9 s+
Claude Haiku 5.5 (no reasoning) median 1.8 s, 1 in 10 moves 2.2 s+
Claude Opus 5.5 median 3.0 s, 1 in 10 moves 3.5 s+
Claude Opus 5.5 (low reasoning) median 2.8 s, 1 in 10 moves 3.3 s+
Claude Sonnet 5.5 median 2.1 s, 1 in 10 moves 2.7 s+
DeepSeek V4 Pro 0423 median 4.6 s, 1 in 10 moves 6.6 s+
DeepSeek V4.1 Flash (low reasoning) median 0.8 s, 1 in 10 moves 1.7 s+
Gemini 3.1 Pro Preview median 2.7 s, 1 in 10 moves 3.4 s+
Gemini 3.1 Pro Preview (low reasoning) median 2.5 s, 1 in 10 moves 3.0 s+
Gemini 3.8 Flash median 2.3 s, 1 in 10 moves 2.9 s+
Muse Spark 1.3 median 4.6 s, 1 in 10 moves 9.5 s+
Muse Spark 1.3 (minimal reasoning) median 2.2 s, 1 in 10 moves 4.9 s+
Muse Wonder median 3.2 s
Muse Wonder (minimal reasoning) median 1.6 s
Kimi K3 median 1.4 s, 1 in 10 moves 6.5 s+
Kimi K3 (low reasoning) median 1.7 s, 1 in 10 moves 4.2 s+
GPT-6 Astra median 2.0 s, 1 in 10 moves 3.4 s+
GPT-6 Astra (low reasoning) median 2.1 s, 1 in 10 moves 3.1 s+
GPT-6 Luna median 2.9 s, 1 in 10 moves 4.2 s+
GPT-6 Luna (no reasoning) median 2.1 s, 1 in 10 moves 3.1 s+
GPT-6 Sol median 1.6 s, 1 in 10 moves 3.7 s+
GPT-6 Sol (no reasoning) median 1.6 s, 1 in 10 moves 2.0 s+
GPT-6.1 Sol median 2.0 s, 1 in 10 moves 3.4 s+
GPT-6.1 Sol (low reasoning) median 1.6 s, 1 in 10 moves 3.2 s+
Qwen3.8 Flash median 2.5 s, 1 in 10 moves 3.8 s+
Qwen3.8 Max (0902) median 3.1 s, 1 in 10 moves 4.6 s+
Qwen3.8 Max (0902) (minimal reasoning) median 2.9 s, 1 in 10 moves 4.9 s+
Grok 4.7 median 2.5 s, 1 in 10 moves 5.5 s+
Grok 4.7 (low reasoning) median 2.3 s, 1 in 10 moves 4.3 s+
MiMo-V2.6-Flash median 3.1 s, 1 in 10 moves 4.6 s+
GLM 4.7 Flash median 4.4 s, 1 in 10 moves 6.0 s+
GLM 5.3 median 1.8 s, 1 in 10 moves 6.2 s+
GLM 5.3 (low reasoning) median 1.2 s, 1 in 10 moves 4.0 s+
GLM 5.3 Flash (low reasoning) median 0.9 s, 1 in 10 moves 2.6 s+

Blitz 3+2

DeepSeek V4 Pro 0423 median 6.4 s, 1 in 10 moves 29.7 s+
Muse Spark 1.3 median 19.4 s, 1 in 10 moves 60.0 s+
Mistral Large 4 (high reasoning) median 26.1 s, 1 in 10 moves 69.5 s+
Kimi K3 median 6.8 s, 1 in 10 moves 32.1 s+
GPT-6 Sol median 8.2 s, 1 in 10 moves 22.9 s+
Qwen3.8 Max (0902) median 7.4 s, 1 in 10 moves 88.9 s+
Grok 4.7 median 9.9 s, 1 in 10 moves 32.9 s+

Decision models

Decision models answer a typed question instead of writing text. Each move is one question whose options are the legal moves, so they cannot play an illegal move, and they are not told their clock. They play the same clock, ladder and engine and are in the table above; here they are ranked among themselves.

No decision model was rated on this measure.

Too slow for the clock. These lost their first four games on time and are not rated for that clock.

Bullet 60s

Kev 4B median 1.9 s, 1 in 10 moves 15.4 s+

Real games: the clock at Lightning 10+1

Each model's clock after every move of one game. A fast model earns back more than it spends; a slow one runs out within a few moves, whatever the position.

Gemini 3.5 Flash LiteWon by checkmate after 28 moves

e40.6 s10.4 s left
Nf30.9 s10.4 s left
Bc40.6 s10.9 s left
O-O0.9 s10.9 s left
c30.6 s11.4 s left
d40.9 s11.4 s left
Nxe51.0 s11.4 s left
Nxc60.6 s11.8 s left
Qxg40.9 s11.9 s left
Qxc8+0.9 s12.0 s left
… 18 more moves

Claude Opus 5.5Lost on time after 3 moves

e43.5 s7.5 s left
Nf33.5 s5.0 s left
d43.3 s2.6 s left

More from the results

Bullet: strength against thinking time

The same at 60 seconds for the whole game, no increment. A 40-move game leaves 1.5 seconds a move.

Named: the six bestOther models (hover for names)

How games are lost

Share of games lost on time, and moves with no legal reply, by clock.

Model	On time, Lightning	On time, Bullet	Invalid moves, Lightning	Invalid moves, Bullet
Gemini 3.5 Flash-LiteGoogle	0.0%	0.0%	0.0%	0.0%
Gemini 3.1 Flash-Lite (minimal reasoning)Google	0.0%	8.3%	0.2%	0.0%
Gemini 3.5 Flash (minimal reasoning)Google	33.3%	25.0%	0.6%	0.5%
Mercury DecideInception	0.0%	0.0%	0.0%	0.0%
GPT-6.1 Sol Ultrafast (low reasoning)OpenAI	50.0%	25.0%	0.0%	0.0%
d1Liquid AI	0.0%	0.0%	0.0%	0.0%
Seed-2.0-Mini (minimal reasoning)ByteDance	0.0%	0.0%	0.8%	0.4%
Mistral Small 4 (no reasoning)Mistral AI	0.0%	25.0%	1.0%	0.6%
GPT-5.4 mini (no reasoning)OpenAI	0.0%	50.0%	0.0%	0.0%
Inkling Small (no reasoning)Thinking Machines	0.0%	0.0%	5.3%	7.2%
Jev 1.13TypeSafe	0.0%	0.0%	0.0%	0.0%
Clef FlashCloudflare	0.0%	0.0%	0.0%	0.0%
ClefCloudflare	0.0%	0.0%	0.0%	0.0%
Mistral Medium 3.5 (no reasoning)Mistral AI	0.0%	0.0%	0.0%	0.4%
Mistral Medium 3.5Mistral AI	0.0%	8.3%	0.3%	0.8%
GPT-6 Luna DecisionsOpenAI	0.0%	0.0%	0.0%	0.0%
Decider V1 27BPerplexity	0.0%	8.3%	0.0%	0.0%
Mistral Large 4Mistral AI	58.3%	–	1.8%	–
Gemini 3.8 Flash (low reasoning)Google	83.3%	50.0%	0.0%	0.0%
GPT-5.4 nano (no reasoning)OpenAI	75.0%	66.7%	4.0%	3.1%
GPT-6.1 Sol UltrafastOpenAI	83.3%	41.7%	0.0%	0.0%
Claude Haiku 4.5Anthropic	83.3%	25.0%	0.0%	0.3%
Claude Sonnet 5.5 (low reasoning)Anthropic	91.7%	–	0.0%	–
Gemini 3.6 Flash (minimal reasoning)Google	91.7%	41.7%	0.0%	0.3%
Kev 4BJared Palmer	66.7%	–	0.0%	–

Lightning: strength against cost

What one Lightning game cost at the run date's prices, on a log scale. Up and to the left is better; the dashed line joins the strongest model at each cost.

Named: the six best and the best for the moneyOther models (hover for names)Best score at each cost

Best per budget

d1: 572 at $0.0005
Gemini 3.1 Flash-Lite (minimal reasoning): 780 at $0.0058
Gemini 3.5 Flash-Lite: 892 at $0.0070

Cheapest first: each model here beats every cheaper one on score.

How BulletBench works

1
The position
The model gets the position, the moves so far, its remaining clock and the legal moves.
2
The reply
It answers with one move. The whole API round trip, including any hidden reasoning, comes off its clock.
3
The engine
Stockfish replies instantly at the model's current ladder level.
4
The ladder
Win and the next game is a level up; lose and it is a level down; draw and it stays. The results give a rating with a 95% interval.

Why chess on a clock

Many products need an answer in about a second: routing, triage, autocomplete, live agents. BulletBench asks which models can still think usefully at that speed. Chess gives a hard, objective score, and the clock makes slow thinking a real cost instead of a free extra.

The clocks

Lightning 10+1: Ten seconds, plus one second a move. A model can play for ever, but only at about a second an answer.
Bullet 60s: One minute for the whole game, no increment.
Blitz 3+2: Three minutes plus two seconds a move: room for reasoning models, for comparison.

Fair play

Same prompt: Every model gets the same system prompt and position format.
Invalid replies: A reply with no legal move gets two corrective retries, then a random legal move is played and counted.
Settings: Each model runs at its provider's default and, where the provider offers one, its fastest reasoning setting, shown separately.

Honest caveats

A dozen games or fewer per clock (the table gives each configuration's count) give wide intervals, so neighbouring ranks are often within each other's ranges. Ratings also come in steps: the engine plays at fixed levels and the rating is fitted from which levels a model beat, drew or lost to, so models with the same results against the same levels get exactly the same rating. Latency depends on the provider's servers on the day. The ladder's Elo anchors come from v1's Stockfish; this edition runs Stockfish 19, so ratings are not comparable with v1's.

What it measures

Decision quality under real time pressure
Response latency at the provider's default settings
Which reasoning settings are fast enough to use

What it does not measure

Any business skill: this is chess against an engine
Chess strength without a clock

Method

Games against a Stockfish ladder; every second of response time comes off the model's clock
Each model at its default setting and, where offered, its fastest reasoning setting
Ladder Elo with a 95% interval; it orders models, it is not a FIDE rating

BulletBench

Ladder Elo: GPT-6.1 Sol

Lightning: strength against thinking time

Full results

Progress over time

Best score so far

Best score by lab

Too slow for the clock

Bullet 60s

Lightning 10+1

Blitz 3+2

Decision models

Bullet 60s

Real games: the clock at Lightning 10+1

More from the results

Bullet: strength against thinking time

How games are lost

Lightning: strength against cost

Best per budget

How BulletBench works

The position

The reply

The engine

The ladder

Why chess on a clock

The clocks

Fair play

Honest caveats

What it measures

What it does not measure

Method