[ BENCHMARKS ]
Fastest on every public benchmark
Six browser benchmarks and the open sandbox lifecycle benchmark, each run on the benchmark’s own public harness.
[
Overview
]
Number one on every benchmark
These are public benchmarks built and published by other teams, not by us. Each one runs the same test against every provider. We ran Solari on the same open-source harness and compare it with the results each board has published.
4×
Faster browser sessions
Than Kernel, the second-fastest provider, on Browserbench.
2×
Browser actions per second
Against Tilion, the next fastest, on BrowserArena.
These speeds are available only on the Enterprise plan.
[
Browsers
]
Fastest to start, fastest to act
Two things decide how fast a browser agent feels. Lifecycle benchmarks time how long it takes to get a browser and hand it back: create a session, connect, load a page, release. Throughput benchmarks time how fast the agent can work once it is inside. Pick a benchmark to see every provider.
Browserbench
4.0×
faster than the next provider
A full session (create, connect, load a page, release) averages 199.9 ms on Solari. Kernel, the next fastest, averages 793.8 ms.
What it measures
The whole life of one browser session, the way an agent uses it: create the session, connect over CDP, load a page, release. Run 5,000 times per provider.
Source: Steel Browserbench ↗Full results
End-to-End Latency
| Provider | Success | Average (ms) | Median (ms) | p95 (ms) | p99 (ms) |
|---|---|---|---|---|---|
| Solari | 100% | 199.90 | 191.00 | 247.00 | 346.05 |
| Kernel | 100% | 793.84 | 776.00 | 1,006.00 | 1,105.00 |
| Steel | 100% | 894.13 | 867.00 | 1,090.00 | 1,340.05 |
| Browserbase | 100% | 2,966.87 | 2,888.00 | 3,886.00 | 4,309.12 |
| Hyperbrowser | 100% | 3,657.11 | 3,665.50 | 5,338.00 | 6,695.05 |
| AnchorBrowser | 97.34% | 8,001.29 | 7,919.00 | 11,561.00 | 13,957.14 |
Browserbench is Steel's open-source benchmark. Solari measured on the same harness; competitor figures are the board's published results as of November 10, 2026.
Browserbench: time per stage
| Provider | Create (ms) | Connect (ms) | Goto (ms) | Release (ms) | Create + connect + release (ms) |
|---|---|---|---|---|---|
| Solari | 9.80 | 25.28 | 161.53 | 3.28 | 38.36 |
| Kernel | 36.64 | 288.30 | 434.15 | 34.77 | 359.71 |
| Steel | 181.57 | 174.64 | 490.29 | 47.62 | 403.83 |
| Browserbase | 212.83 | 1,794.42 | 745.08 | 214.55 | 2,221.80 |
| Hyperbrowser | 1,731.63 | 347.60 | 377.89 | 1,199.98 | 3,279.21 |
| AnchorBrowser | 3,796.55 | 184.29 | 1,259.66 | 2,760.79 | 6,741.63 |
Average time per stage in ms, 5,000 runs per provider. Page load (goto) depends on the target site, so the last column leaves it out.
ComputeSDK lifecycle
3.6×
faster than the next provider
A full session takes 43.2 ms at the median, with a p99 of 49.2 ms. That p99 is under a third of the runner-up's median.
What it measures
Create, connect, navigate and release a headless browser session. The board scores each provider on its median and on its slowest runs, so one bad tail costs points.
Source: ComputeSDK browser benchmarks ↗Full results
ComputeSDK Browser Lifecycle: end-to-end latency
| Provider | Score | Success | p50 (ms) | p90 (ms) | p99 (ms) |
|---|---|---|---|---|---|
| Solari | 99.67 | 100% | 43.2 | 46.7 | 49.2 |
| Kernel | 98.44 | 100% | 156.2 | 202.8 | 377.6 |
| Browser Use | 97.06 | 100% | 320.5 | 356.4 | 378.6 |
| Tilion | 94.54 | 100% | 205.4 | 1,490.2 | 1,548.6 |
| Hyperbrowser | 94.44 | 100% | 680.9 | 719.0 | 739.8 |
| Browserbase | 94.18 | 100% | 736.6 | 804.7 | 883.6 |
| Notte | 84.46 | 99% | 1,693.5 | 1,882.4 | 2,489.0 |
| Steel | 78.41 | 99% | 2,533.0 | 3,056.5 | 3,434.7 |
Score is ComputeSDK's composite (0.60 × p50 + 0.25 × p95 + 0.15 × p99 scores, × success rate). Competitor figures: ComputeSDK board run of November 6, 2026.
ComputeSDK Browser Lifecycle: time per stage
| Provider | Create (ms) | Connect (ms) | Navigate (ms) | Release (ms) |
|---|---|---|---|---|
| Solari | 3.9 | 13.0 | 23.0 | 1.3 |
| Kernel | 8.8 | 26.4 | 89.1 | 15.2 |
| Browser Use | 183.4 | 37.6 | 62.2 | 23.4 |
| Tilion | 12.9 | 116.1 | 60.3 | 13.7 |
| Hyperbrowser | 216.6 | 154.2 | 207.7 | 89.9 |
| Browserbase | 115.4 | 299.0 | 197.7 | 108.7 |
| Notte | 378.1 | 416.3 | 731.2 | 97.5 |
| Steel | 246.6 | 1,902.6 | 134.7 | 158.2 |
Median (p50) time per stage in ms. Competitor figures: ComputeSDK board run of November 6, 2026.
BrowserArena, 1 session
2.7×
faster than the next provider
59 ms at the median with one session at a time. Tilion, the next fastest, takes 159 ms.
What it measures
The same create, connect, load and release flow, with one session running at a time.
Source: BrowserArena ↗Full results
BrowserArena hello-browser: 1 session
| Provider | Success | p50 (ms) | p90 (ms) | p95 (ms) | p99 (ms) |
|---|---|---|---|---|---|
| Solari | 100% | 59 | 70 | 73 | 85 |
| Tilion | 99% | 159 | 200 | 338 | 13,511 |
| Kernel | 100% | 273 | 410 | 483 | 815 |
| Notte | 100% | 284 | 340 | 359 | 495 |
| Browser Use | 100% | 388 | 430 | 441 | 459 |
| Lightpanda | 100% | 439 | 455 | 460 | 531 |
| TinyFish | 100% | 605 | 664 | 720 | 809 |
| Browserbase | 100% | 785 | 866 | 932 | 972 |
| Cloudflare | 100% | 935 | 1,264 | 1,411 | 1,839 |
| Steel | 97% | 1,161 | 1,329 | 1,466 | 3,099 |
| AWS AgentCore | 100% | 2,444 | 3,152 | 3,294 | 3,396 |
| AnchorBrowser | 100% | 2,883 | 7,472 | 8,942 | 21,827 |
| Hyperbrowser | 0% | — | — | — | — |
Total session time in ms. Competitor figures: BrowserArena board run of November 10, 2026. Hyperbrowser completed no sessions that day.
BrowserArena hello-browser, 1 session: time per stage
Solari
- Create
- 4
- Connect
- 16
- Goto
- 38
- Release
- 1
Tilion
- Create
- 17
- Connect
- 98
- Goto
- 30
- Release
- 14
Kernel
- Create
- 22
- Connect
- 51
- Goto
- 184
- Release
- 16
Notte
- Create
- 136
- Connect
- 25
- Goto
- 86
- Release
- 37
Browser Use
- Create
- 176
- Connect
- 110
- Goto
- 78
- Release
- 25
Lightpanda
- Create
- 0
- Connect
- 294
- Goto
- 124
- Release
- 21
TinyFish
- Create
- 141
- Connect
- 109
- Goto
- 304
- Release
- 52
Browserbase
- Create
- 126
- Connect
- 369
- Goto
- 182
- Release
- 109
Cloudflare
- Create
- 192
- Connect
- 363
- Goto
- 257
- Release
- 124
Steel
- Create
- 220
- Connect
- 667
- Goto
- 121
- Release
- 153
AWS AgentCore
- Create
- 1,367
- Connect
- 205
- Goto
- 772
- Release
- 100
AnchorBrowser
- Create
- 1,896
- Connect
- 134
- Goto
- 133
- Release
- 720
| Provider | Create (ms) | Connect (ms) | Goto (ms) | Release (ms) |
|---|---|---|---|---|
| Solari | 4 | 16 | 38 | 1 |
| Tilion | 17 | 98 | 30 | 14 |
| Kernel | 22 | 51 | 184 | 16 |
| Notte | 136 | 25 | 86 | 37 |
| Browser Use | 176 | 110 | 78 | 25 |
| Lightpanda | 0 | 294 | 124 | 21 |
| TinyFish | 141 | 109 | 304 | 52 |
| Browserbase | 126 | 369 | 182 | 109 |
| Cloudflare | 192 | 363 | 257 | 124 |
| Steel | 220 | 667 | 121 | 153 |
| AWS AgentCore | 1,367 | 205 | 772 | 100 |
| AnchorBrowser | 1,896 | 134 | 133 | 720 |
Median (p50) time per stage in ms. Competitor figures: BrowserArena board run of November 10, 2026.
BrowserArena, 10 sessions
2.4×
faster than the next provider
108 ms at the median with ten sessions running at once, and a p99 of 178 ms. The runner-up's p99 rises above 27 seconds.
What it measures
The same flow with ten sessions running at once. It shows whether a provider stays fast when an agent fleet arrives together.
Source: BrowserArena ↗Full results
BrowserArena hello-browser: 10 concurrent sessions
| Provider | Success | p50 (ms) | p90 (ms) | p95 (ms) | p99 (ms) |
|---|---|---|---|---|---|
| Solari | 100% | 108 | 152 | 162 | 178 |
| Tilion | 99% | 258 | 499 | 11,106 | 27,715 |
| Kernel | 100% | 321 | 513 | 2,469 | 4,440 |
| Notte | 100% | 411 | 680 | 768 | 827 |
| Browser Use | 100% | 530 | 761 | 865 | 1,116 |
| Lightpanda | 100% | 727 | 885 | 907 | 1,482 |
| Browserbase | 100% | 784 | 55,035 | 55,091 | 58,120 |
| Cloudflare | 100% | 967 | 1,334 | 1,520 | 2,820 |
| Steel | 98% | 1,229 | 1,454 | 1,588 | 1,795 |
| AWS AgentCore | 100% | 2,531 | 3,337 | 3,670 | 4,067 |
| AnchorBrowser | 100% | 3,621 | 5,558 | 6,437 | 15,333 |
| Hyperbrowser | 0% | — | — | — | — |
Total session time in ms. Competitor figures: BrowserArena board run of November 10, 2026. Hyperbrowser completed no sessions that day.
BrowserArena hello-browser, 10 sessions: time per stage
Solari
- Create
- 14
- Connect
- 44
- Goto
- 49
- Release
- 1
Tilion
- Create
- 67
- Connect
- 136
- Goto
- 40
- Release
- 15
Kernel
- Create
- 36
- Connect
- 89
- Goto
- 177
- Release
- 19
Notte
- Create
- 251
- Connect
- 29
- Goto
- 91
- Release
- 40
Browser Use
- Create
- 216
- Connect
- 181
- Goto
- 96
- Release
- 37
Lightpanda
- Create
- 0
- Connect
- 582
- Goto
- 125
- Release
- 21
Browserbase
- Create
- 149
- Connect
- 369
- Goto
- 162
- Release
- 104
Cloudflare
- Create
- 227
- Connect
- 359
- Goto
- 247
- Release
- 135
Steel
- Create
- 311
- Connect
- 653
- Goto
- 105
- Release
- 161
AWS AgentCore
- Create
- 1,461
- Connect
- 204
- Goto
- 768
- Release
- 99
AnchorBrowser
- Create
- 2,294
- Connect
- 158
- Goto
- 229
- Release
- 940
| Provider | Create (ms) | Connect (ms) | Goto (ms) | Release (ms) |
|---|---|---|---|---|
| Solari | 14 | 44 | 49 | 1 |
| Tilion | 67 | 136 | 40 | 15 |
| Kernel | 36 | 89 | 177 | 19 |
| Notte | 251 | 29 | 91 | 40 |
| Browser Use | 216 | 181 | 96 | 37 |
| Lightpanda | 0 | 582 | 125 | 21 |
| Browserbase | 149 | 369 | 162 | 104 |
| Cloudflare | 227 | 359 | 247 | 135 |
| Steel | 311 | 653 | 105 | 161 |
| AWS AgentCore | 1,461 | 204 | 768 | 99 |
| AnchorBrowser | 2,294 | 158 | 229 | 940 |
Median (p50) time per stage in ms. Competitor figures: BrowserArena board run of November 10, 2026.
ComputeSDK throughput
9.8/s
browser actions per second
The highest action rate on the board and the top composite score: 94.32 against 87.86 for the runner-up.
What it measures
How fast an agent can work once the browser is open: ten actions in a row, such as navigate, wait for a selector, screenshot, read text, click and go back.
Source: ComputeSDK browser benchmarks ↗Full results
ComputeSDK Browser Throughput
| Provider | Score | Success | Actions/s p50 | Task p50 (ms) | Task p95 (ms) | Task p99 (ms) |
|---|---|---|---|---|---|---|
| Solari | 94.32 | 97% | 9.80 | 1,019 | 1,493 | 1,640 |
| Kernel | 87.86 | 100% | 7.58 | 1,319 | 1,904 | 2,252 |
| Tilion | 80.81 | 100% | 7.66 | 1,305 | 13,005 | 13,995 |
| Browserbase | 77.87 | 100% | 5.32 | 1,880 | 2,652 | 2,857 |
| Steel | 54.41 | 76% | 4.03 | 2,397† | 3,583 | 30,000 |
| Browser Use | 34.87 | 41% | 6.59 | 113† | 1,727 | 1,782 |
| Hyperbrowser | 26.74 | 38% | 3.25 | 225† | 3,398 | 3,816 |
| Notte | 3.30 | 5% | 1.65 | 185† | 209 | 261 |
A task is 10 browser actions: navigate, wait for a selector, screenshot, read text, click, go back. Score is the board's composite, weighted by success rate. Competitor figures: ComputeSDK board run of November 6, 2026. † Includes sessions that failed early.
BrowserArena throughput
2.0×
the actions per second of the next provider
10.23 browser actions per second against 5.08 for the next fastest. A 10-action task finishes in under a second.
What it measures
Ten browser actions per session, counted as actions per second. Start-up time is left out, so this is purely how quickly the browser responds.
Source: BrowserArena ↗Full results
BrowserArena browser-throughput
| Provider | Success | Actions/s p50 | Task p50 (ms) | Task p95 (ms) | Task p99 (ms) |
|---|---|---|---|---|---|
| Solari | 100% | 10.23 | 981 | 1,604 | 1,863 |
| Tilion | 100% | 5.08 | 1,970 | 3,347 | 4,359 |
| Notte | 94% | 4.44 | 2,250 | 60,113 | 60,128 |
| Browser Use | 95% | 4.16 | 2,405 | 3,743 | 60,335 |
| Kernel | 96% | 3.79 | 2,632 | 5,098 | 60,290 |
| Browserbase | 100% | 3.55 | 2,817 | 4,271 | 5,326 |
| Steel | 99% | 3.46 | 2,884 | 4,832 | 5,809 |
| TinyFish | 90% | 3.30 | 3,034 | 60,210 | 60,222 |
| AnchorBrowser | 100% | 3.08 | 3,238 | 5,062 | 6,614 |
| Cloudflare | 100% | 1.71 | 5,828 | 9,217 | 12,364 |
| AWS AgentCore | 71% | 1.08 | 9,316 | 36,761 | 50,289 |
Task times in ms. Competitor figures: BrowserArena board run of November 10, 2026. Hyperbrowser and Lightpanda completed no sessions.
[
Sandboxes
]
Sandbox benchmarks, start to finish
This benchmark does what an agent does with code it just wrote: create a sandbox, run a heavy job, clean up. Every provider runs the same job on the same open-source harness.
Full lifecycle
30%
faster than the next provider
Create a sandbox, run a heavy workload and clean up: 8.2 s on Solari. E2B, the next fastest, takes 10.8 s.
What it measures
What an agent does when it runs code it just wrote: create a sandbox, run a heavy single-threaded job, and clean up. Every provider runs the same job.
Source: nibzard ai-sandbox-benchmark ↗Full results
Full Sandbox Lifecycle
| Provider | Create (ms) | Run code (ms) | Clean up (ms) | Total (ms) |
|---|---|---|---|---|
| Solari | 869 | 7,019 | 269 | 8,158 |
| E2B | 541 | 9,868 | 401 | 10,810 |
| Modal | 2,070 | 6,607 | 3,235 | 11,912 |
| Daytona | 2,202 | 8,531 | 141 | 13,589 |
| CodeSandbox | 1,321 | 17,237 | 6,234 | 24,792 |
The open-source nibzard test_fft_performance lifecycle, same harness for every provider. Competitor figures are the benchmark's own published results. Solari ran at 1 vCPU / 8 GB, with its client a region away from its gateway, so create and clean up are measured conservatively.
Give every agent
a place to work
Browsers, sandboxes and desktops from one API. Start free and scale when you need to.
