Claude Code, Codex & Antigravity today: measured speed, thinking and accuracy
As of · measured every hour from one EU server on the vendors' own plans (Claude Max subscription, ChatGPT Free plan, Google AI subscription) · each agent is compared with its own last 7 days
Claude Code (Claude Opus 5.5) generated 114.8 tokens/s over the last 24 hours vs 115.6 tokens/s in the 7 days before (−1%) — no significant change (p=0.696). Codex CLI (GPT-6 Luna): collection paused since 2026-11-18 09:37 UTC (3 hours ago): our budget guard stopped the calls because the account's usage limit is nearly used up; collection resumes by itself when the window resets. Antigravity CLI (Gemini 3.8 Flash) generated 156.2 tokens/s over the last 24 hours vs 154.6 tokens/s in the 7 days before (+1%) — no significant change (p=0.98). Claude Code (Claude Sonnet 5.5) generated 148.5 tokens/s over the last 24 hours vs 146.0 tokens/s in the 7 days before (+2%) — no significant change (p=0.0779). Antigravity CLI (GPT-OSS 120B) generated 151.7 tokens/s over the last 24 hours vs 161.1 tokens/s in the 7 days before (−6%) — no significant change (p=0.314).
In this section: Codex vs Claude Code, side by side · Is Claude Code down or slower today? · Is Codex down or slower today? · Is Antigravity CLI down or slower today?
Is a coding agent slower or "dumber" today, or does it just feel that way? We run the same fixed task through Claude Code, Codex CLI and Antigravity CLI every hour and record how fast the answer arrives, how many tokens the model spends thinking, and how long the first token takes. A fixed hard question panel, asked in slices, checks accuracy. These are measurements of what a user of that plan actually gets, not vendor claims.
Right now
Claude Code · Claude Opus 5.5: no significant change in the last 24 hours
Last measured 2026-11-18 12:37 UTC · claude-cli 2.1.327 (Claude Code) · effort high · Claude Max subscription · errors in 24 h: 0% · served model claude-opus-5-5 as reported by the CLI (0 mismatches in 7 days)
| Metric | Last 24 h | 7 days before | Change | Verdict |
|---|---|---|---|---|
| Speed | 114.8 tokens/s | 115.6 | −1% | no significant change (p=0.696) |
| Thinking | 842 tokens | 776 | +9% | no significant change (p=0.263) |
| First token | 2.4 s | 2.3 s | +5% | no significant change (p=0.118) |
Codex CLI · GPT-6 Luna: collection paused since 2026-11-18 09:37 UTC
Our budget guard stopped the calls because the account's usage limit is nearly used up; collection resumes by itself when the window resets. Last measured 2026-11-18 09:37 UTC · codex 0.156.1 · effort high · ChatGPT Free plan · errors in 24 h: 0% · requested gpt-6-luna; the CLI does not report which model answered, so a silent substitution would not be visible · measured on the ChatGPT Free plan with its default model since 2026-11-11; GPT-6 Astra on the Pro plan was measured until 2026-10-22 and its history is kept apart from this series
| Metric | Last 24 h | 7 days before | Change | Verdict |
|---|---|---|---|---|
| Speed | 49.9 tokens/s | 50.0 | ±0% | paused |
| Thinking | 792 tokens | 813 | −3% | paused |
| First token | 11.1 s | 11.1 s | ±0% | paused |
Antigravity CLI · Gemini 3.8 Flash: no significant change in the last 24 hours
Last measured 2026-11-18 12:39 UTC · agy 1.5.11 · effort high · Google AI subscription · errors in 24 h: 0% · requested gemini-3.8-flash-high; the CLI does not report which model answered, so a silent substitution would not be visible · thinking is not streamed by the CLI, so the first token is the first visible token after thinking; tools are denied by permission policy, a tool attempt counts as an error; the hard panel is asked in one slice a day (15-16 questions after 00:47 UTC, the whole panel every 5 days) because Flash spends 5-45k tokens per question
| Metric | Last 24 h | 7 days before | Change | Verdict |
|---|---|---|---|---|
| Speed | 156.2 tokens/s | 154.6 | +1% | no significant change (p=0.98) |
| Thinking | 14,667 tokens | 14,788 | −1% | no significant change (p=0.619) |
| First token | 97.8 s | 95.8 s | +2% | no significant change (p=0.7) |
Claude Code · Claude Sonnet 5.5: no significant change in the last 24 hours
Last measured 2026-11-18 12:37 UTC · claude-cli 2.1.327 (Claude Code) · effort high · Claude Max subscription · errors in 24 h: 0% · served model claude-sonnet-5-5 as reported by the CLI (0 mismatches in 7 days)
| Metric | Last 24 h | 7 days before | Change | Verdict |
|---|---|---|---|---|
| Speed | 148.5 tokens/s | 146.0 | +2% | no significant change (p=0.0779) |
| Thinking | 652 tokens | 639 | +2% | no significant change (p=0.47) |
| First token | 2.2 s | 2.2 s | ±0% | no significant change (p=0.787) |
Antigravity CLI · GPT-OSS 120B: no significant change in the last 24 hours
Last measured 2026-11-18 12:37 UTC · agy 1.5.11 · effort medium · Google AI subscription · errors in 24 h: 0% · requested gpt-oss-120b-medium; the CLI does not report which model answered, so a silent substitution would not be visible · an open-weight model served by Google inside Antigravity CLI on the same plan as Gemini 3.8 Flash; the CLI reports no thinking tokens for it (it does reason: the first visible token comes after the reasoning), so thinking is not measured; tools are denied by permission policy; the hard panel is asked in one slice a day (39 questions after 00:47 UTC, the whole panel every 2 days)
| Metric | Last 24 h | 7 days before | Change | Verdict |
|---|---|---|---|---|
| Speed | 151.7 tokens/s | 161.1 | −6% | no significant change (p=0.314) |
| Thinking | — tokens | — | — | not reported by the CLI |
| First token | 11.2 s | 11.3 s | ±0% | no significant change (p=0.932) |
Speed is output tokens (thinking included) per second of response time for one fixed task with a single exact answer. A change counts as significant only when the median moves by at least 15% and a Mann-Whitney test gives p < 0.01, with at least 12 recent and 48 baseline samples. Do not compare these numbers between agents: Claude Code reports API time, Codex and Antigravity report turn time, Antigravity does not stream thinking, and their tokenizers differ. What is comparable — time to answer the same task, failed calls and accuracy on the same questions — is on Codex vs Claude Code, side by side.
Hourly speed, last 7 days
Claude Code · Claude Opus 5.5
Codex CLI · GPT-6 Luna
Antigravity CLI · Gemini 3.8 Flash
Claude Code · Claude Sonnet 5.5
Antigravity CLI · GPT-OSS 120B
Daily medians
| UTC day | Claude Opus 5.5 tokens/s | thinking | GPT-6 Luna tokens/s | thinking | Gemini 3.8 Flash tokens/s | thinking | Claude Sonnet 5.5 tokens/s | thinking | GPT-OSS 120B tokens/s | thinking |
|---|---|---|---|---|---|---|---|---|---|---|
| 2026-11-18 | 115.5 | 944 | 50.4 | 757 | 156.7 | 14,671 | 149.0 | 638 | 141.6 | — |
| 2026-11-17 | 115.0 | 720 | 49.6 | 795 | 151.8 | 14,521 | 146.1 | 660 | 167.3 | — |
| 2026-11-16 | 116.2 | 730 | 49.7 | 739 | 158.4 | 15,397 | 148.6 | 654 | 159.9 | — |
| 2026-11-15 | 114.2 | 780 | 50.4 | 812 | 163.8 | 14,822 | 145.0 | 642 | 165.4 | — |
| 2026-11-14 | 114.1 | 873 | 49.4 | 747 | 155.4 | 14,727 | 141.6 | 639 | 167.2 | — |
| 2026-11-13 | 115.9 | 890 | 49.9 | 817 | 153.4 | 14,772 | 149.5 | 634 | 156.3 | — |
| 2026-11-12 | 115.0 | 710 | 50.4 | 856 | 155.9 | 15,306 | 143.9 | 616 | 144.7 | — |
| 2026-11-11 | 115.9 | 846 | 49.4 | 902 | 150.8 | 14,182 | 145.3 | 636 | 173.1 | — |
| 2026-11-10 | 115.2 | 777 | — | — | 153.0 | 14,212 | — | — | — | — |
| 2026-11-09 | 114.9 | 898 | — | — | 155.4 | 15,388 | — | — | — | — |
The CLI version is recorded with every measurement, so a regression in the tool can be told apart from a change in the model behind it. Claude Code CLI changed 11 times (claude-cli 2.1.309 (Claude Code) → claude-cli 2.1.326 (Claude Code)); Codex CLI CLI stayed at codex 0.156.1; Antigravity CLI CLI changed 7 times (agy 1.5.0 → agy 1.5.10).
Accuracy on a hard question panel
Easy questions cannot show a weaker model: Claude Code and Codex solved 99–100% of our earlier synthetic tasks. The hard panel has 78 questions (59 MMLU-Pro, 12 GPQA Diamond, 4 olympiad math (MathArena), 3 AIME 2026) that Claude Opus 5.5 solves only sometimes — 62% in the calibration run of livenerf, whose panel we reuse unchanged (MIT licence). Each question is asked once a day, or in daily slices for models whose thinking is expensive (the schedule is under How it is measured); the questions themselves are not published.
| Agent | Answered | Correct | Accuracy | Thinking (median) | By UTC day |
|---|---|---|---|---|---|
| Claude Code · Claude Opus 5.5 | 2884 | 1747 | 61% | 376 | 11-12: 47/78 · 11-13: 53/78 · 11-14: 48/78 · 11-15: 49/78 · 11-16: 51/78 · 11-17: 51/78 · 11-18: 24/39 |
| Codex CLI · GPT-6 Luna | 544 | 234 | 43% | 306 | 11-12: 33/78 · 11-13: 35/78 · 11-14: 33/78 · 11-15: 37/78 · 11-16: 32/78 · 11-17: 30/75 · 11-18: 17/39 |
| Antigravity CLI · Gemini 3.8 Flash | 217 | 116 | 53% | 7,020 | 11-12: 8/15 · 11-13: 1/1 · 11-14: 7/16 · 11-15: 9/16 · 11-16: 6/15 · 11-17: 9/15 · 11-18: 10/16 |
| Claude Code · Claude Sonnet 5.5 | 547 | 267 | 49% | 416 | 11-12: 38/78 · 11-13: 39/78 · 11-14: 40/78 · 11-15: 37/78 · 11-16: 36/78 · 11-17: 39/78 · 11-18: 17/39 |
| Antigravity CLI · GPT-OSS 120B | 306 | 120 | 39% | — | 11-12: 12/37 · 11-13: 13/38 · 11-14: 18/39 · 11-15: 13/38 · 11-16: 16/39 · 11-17: 15/39 · 11-18: 16/38 |
- Claude Code (Claude Opus 5.5): last 7 days vs baseline +3.4 points ± 3.0 (z=1.15, 78 questions) — no significant change.
- Codex CLI (GPT-6 Luna): 234/544 correct (43%) before the pause; the hard panel stopped with the rest of the collection on 2026-11-18.
- Antigravity CLI (Gemini 3.8 Flash): last 7 days vs baseline +0.8 points ± 2.9 (z=0.28, 62 questions) — no significant change.
- Claude Code (Claude Sonnet 5.5): 267/547 correct (49%) so far; the baseline is still being collected (day 7 of 10), so no verdict until 2026-11-21.
- Antigravity CLI (GPT-OSS 120B): 120/306 correct (39%) so far; the baseline is still being collected (day 7 of 10), so no verdict until 2026-11-21.
Rule: baseline = first 10 days of the hard panel, then the last 7 days item by item (paired difference, item-clustered SE); significant when |delta| >= 5 points and |z| >= 2.58. Only Claude Opus 5.5 was used to select the questions (ones it solves only sometimes), so the panel is not a neutral ranking of models: a drop of GPT-6 Luna, Gemini 3.8 Flash, Claude Sonnet 5.5 and GPT-OSS 120B is judged only against each model's own history. Side by side, on the questions both agents answered in the same week: Codex vs Claude Code.
Want other agents measured?
We track 5 agents today. If more people want a model tracked, we add it.
How it is measured
- What
- Claude Code with Claude Opus 5.5 (effort high); Codex CLI with GPT-6 Luna (effort high); Antigravity CLI with Gemini 3.8 Flash (effort high); Claude Code with Claude Sonnet 5.5 (effort high); Antigravity CLI with GPT-OSS 120B (effort medium); tools disabled or denied by permission policy (a tool attempt is recorded as an error, not a measurement), a fresh isolated session per call, on the plans named in each card (Claude Max subscription, ChatGPT Free plan, Google AI subscription).
- When
- Speed: every hour at :37 UTC, one fixed task per agent. Hard panel: every 6 hours (00:47, 06:47, 12:47, 18:47 UTC), a quarter of the panel; each question is asked once a day. Antigravity CLI: one slice a day (the first run after 00:47 UTC, in panel order): Gemini 3.8 Flash 15-16 questions, the whole panel every 5 days; GPT-OSS 120B 39 questions, the whole panel every 2 days.
- Where
- one server in the EU (Hetzner, Germany); a single account per vendor. A slowdown on our account is a measured fact about that account, not proof that every user is affected.
- Errors
- a failed call is 'not measured', never a wrong answer.
- Limits
- We measure delivered behaviour of the subscription — rate limits, routing and defaults included — not the model weights. Codex and Antigravity do not report which model served an answer; Antigravity does not stream thinking, so its first token is measured after the thinking phase.
- Data
- /api/coding-agents.json (JSON, CC BY 4.0) · markdown · site methodology