How fast is the Claude API? Opus 5 time to first token from 5 regions, measured daily
We ran real, authenticated streaming requests against Anthropic's claude-opus-5 with effort set to "low", from Amsterdam, San Francisco, Montreal, Singapore and Tokyo, every day from 29 August to 4 September 2026, plus two controlled runs of 20 requests per city. We parsed the event stream ourselves and timestamped the first visible token: the time to first token (TTFT).
Quick take
Claude's first token typically arrived in 1.37 to 1.51 seconds, measured 35 times per city over seven days. San Francisco was quickest at 1369 ms and Singapore slowest at 1511 ms, a 10% spread, so where you call from changes little. All 200 completions in the two controlled runs succeeded, but the slower responses moved between runs: 1.3 to 1.5 seconds on 30 August, 2.1 to 2.3 seconds on 4 September. That puts Claude about 0.4 seconds behind the ChatGPT API and about 1.2 seconds ahead of Grok in the same daily collection, on typical and on slow responses alike.
- • Typical time to first token: 1369 ms (San Francisco) to 1511 ms (Singapore), pooled from 35 daily requests per city
- • 200 of 200 controlled completions succeeded across two runs; the slower responses were 1.3–1.5 s on 30 August and 2.1–2.3 s on 4 September
- • Day to day, a city's median moved between 1116 and 1601 ms, a wider range than the gap between cities
- • Same prompt, same low reasoning effort, same days: ChatGPT about 1.0 s, Claude about 1.4 s, Grok about 2.6 s
Where in the world is it fast?
San Francisco came out quickest at 1369 ms, then Montreal (1390 ms), Tokyo (1447 ms), Amsterdam (1448 ms) and Singapore (1511 ms). Every city reached the same IP address and finished the TLS handshake within 6 to 26 ms, so the network explains only a small part of the wait. Most of each second and a half is spent on Anthropic's side: queueing plus the low-effort thinking that runs before the first word.
Amsterdam1448 msslower 1777 ms
San Francisco1369 msslower 1871 ms
Montreal1390 msslower 1734 ms
Singapore1511 msslower 1888 ms
Tokyo1447 msslower 1860 ms
Where is the time actually going?
The stream opened at the same moment the first visible token arrived, within 10 ms in every city, and the whole reply to our one-word prompt was done 11 to 37 ms later. Anthropic did not stream a thinking phase in these runs, so the wait shows up as one block before the stream opens rather than as visible reasoning inside it. Connection setup took 2 to 23 ms and TLS 6 to 26 ms.
Click a city to see its stage-by-stage breakdown.
Amsterdam1255 ms
Amsterdam- Finding the server
- 1 ms
- Reaching the server
- 9 ms
- Setting up security
- 12 ms
- Waiting for the server: most of the time
- 1223 ms
- Receiving the response
- 11 ms
- Total
- 1255 ms
- Response opened
- 1244 ms
- First visible word
- 1244 ms
- Answer complete
- 1255 ms
San Francisco1139 ms
San Francisco- Finding the server
- 12 ms
- Reaching the server
- 4 ms
- Setting up security
- 6 ms
- Waiting for the server: most of the time
- 1087 ms
- Receiving the response
- 42 ms
- Total
- 1139 ms
- Response opened
- 1104 ms
- First visible word
- 1114 ms
- Answer complete
- 1139 ms
Montreal1214 ms
Montreal- Finding the server
- 1 ms
- Reaching the server
- 23 ms
- Setting up security
- 26 ms
- Waiting for the server: most of the time
- 1153 ms
- Receiving the response
- 12 ms
- Total
- 1214 ms
- Response opened
- 1202 ms
- First visible word
- 1202 ms
- Answer complete
- 1214 ms
Singapore1314 ms
Singapore- Finding the server
- 3 ms
- Reaching the server
- 19 ms
- Setting up security
- 23 ms
- Waiting for the server: most of the time
- 1261 ms
- Receiving the response
- 11 ms
- Total
- 1314 ms
- Response opened
- 1300 ms
- First visible word
- 1300 ms
- Answer complete
- 1314 ms
Tokyo1262 ms
Tokyo- Finding the server
- 1 ms
- Reaching the server
- 2 ms
- Setting up security
- 9 ms
- Waiting for the server: most of the time
- 1239 ms
- Receiving the response
- 12 ms
- Total
- 1262 ms
- Response opened
- 1225 ms
- First visible word
- 1225 ms
- Answer complete
- 1262 ms
The bars and phases below come from the controlled run of 4 September 2026 (20 requests per city), so they add up within that run. The typical and slower times in the table above pool every daily run, which is why the two sets of numbers differ.
Finding the server is measured once per location, so it sits outside these bars. Open a city to see it.
Bars show the typical request, so the parts add up to its total. Why
Every city, side by side
One row per city, with the typical time highlighted. For the stage-by-stage detail behind any city, open it in the chart above.
| Test location | Typical first word | On a bad day | Slowest seen |
|---|---|---|---|
Amsterdam | 1448 ms | 1777 ms | 2223 ms |
San Francisco Fastest | 1369 ms | 1871 ms | 2315 ms |
Montreal | 1390 ms | 1734 ms | 2206 ms |
Singapore Slowest | 1511 ms | 1888 ms | 2312 ms |
Tokyo | 1447 ms | 1860 ms | 2299 ms |
Global presence
All five cities resolved api.anthropic.com to one IP address, 160.79.104.10, which our verdict labels single origin. The certificate is issued by Google Trust Services and the TLS session ended at a point of presence near each city (London for Amsterdam, San Jose for San Francisco, Chicago for Montreal, Singapore and Tokyo locally), which suggests a front door close to the caller rather than one machine. For a streaming API the delivery verdict comes from stream-open times, and here stream-open includes the model's thinking, so read the 196 ms spread as an upper bound on what the network adds.
All 5 regions resolved to one IP address and latency varies by 196ms across regions. Traffic appears to reach a single origin server.
Conclusions
- • Claude sits between the other two flagship APIs we measure the same way: about 0.4 s behind ChatGPT's first token and about 1.2 s ahead of Grok's, on the same prompt at low reasoning effort during the same days. The slower responses line up the same way: 1.7 to 1.9 s for Claude, 1.1 to 1.9 s for ChatGPT, 3.3 to 3.9 s for Grok.
- • The slow tail moves between days more than between cities. The 95th percentile was 1335 to 1472 ms on 30 August and 2132 to 2303 ms on 4 September, while the typical first token barely changed. The table shows the median of the two runs.
- • Geography barely matters: a 10% spread between the fastest and slowest city, against a spread of up to 40% within one city across seven days of daily medians.
- • Anthropic holds the response until the first token, like OpenAI and unlike xAI: stream open and first token were within 10 ms of each other everywhere.
Plan for 1.4 seconds typical and 2.3 on a slow day
Claude's first token is slower than ChatGPT's and faster than Grok's, and it hardly depends on the city. What changes is the day. A wait of about 1.4 seconds before the first word was the normal case everywhere we tested, while the slower responses ran 1.5 seconds on one controlled run and 2.3 seconds on another, and a city's daily median moved by as much as 0.4 seconds between runs. Budget for that day-to-day drift, not for a closer region, and keep streaming so the wait is visible progress rather than a frozen screen.
How this test was run
7 daily quick runs from 29 August 2026 to 4 September 2026 (35 requests per city), plus 2 controlled runs of 20 requests per city, the latest on 4 September 2026 · 29 August 2026 – 4 September 2026
Full test parameters
- Request
- Authenticated streaming POST to api.anthropic.com/v1/messages
- Measured from
- Amsterdam · San Francisco · Montreal · Singapore · Tokyo
- Model
- claude-opus-5 · output_config effort "low" (adaptive thinking on, not streamed) · max_tokens 256 · anthropic-version 2023-06-01
- Prompt
- "Say 'ok' and nothing else." Identical across every city and every LLM API we benchmark
- Comparison figures
- ChatGPT (gpt-5.6-sol, reasoning_effort low) and Grok (grok-4.6, reasoning_effort low) come from the same daily collection between 27 August and 4 September 2026
- Timings taken
- Parsed from the token stream: stream open, first visible token, completion, plus DNS, connect and TLS
- Typical response time
- Median of all 35 daily requests per city between 29 August 2026 and 4 September 2026; not an average of daily medians
- Slower response time
- Median of the 95th percentiles of the 2 most recent controlled 20-request runs per city
- Day-to-day range (daily medians)
- Amsterdam 1223–1510 ms · San Francisco 1116–1517 ms · Montreal 1163–1549 ms · Singapore 1299–1601 ms · Tokyo 1214–1505 ms
Independent measurement by LatencyRadar. Not affiliated with Anthropic. · How we measure (v1.0)
Get the monthly AI latency report, by test location
One email a month: how fast the major AI providers respond, measured from 5 regions.
Check your inbox for the confirmation email. nothing is sent until you confirm. Unsubscribe any time.
How fast is your API compared to Anthropic's?
Run a free speed test from multiple cities and find out where your users are waiting. No setup, no account required.
Takes about 30 seconds.