Realtime benchmark / 2026-11-09
Codex · Claude Code · DeepSeek Harness Custom skills supplied: 17
Main prompt
Build a global todo app with one shared list that everyone can view, add to, and mark complete without signing up. Generate 100 varied, fun, and achievable todo suggestions. Users can only add todos by choosing from these suggestions; they cannot write their own. Keep at most 200 todos in the shared list. When adding a todo would exceed this limit, automatically remove the oldest todo by creation time, whether completed or not.
One independent run per configuration, using the same exact task. Initial workspace cleanliness and full context isolation were not proved. Charts exclude setup, review, and publication work. Token costs are estimates and exclude infrastructure. Claude includes a Haiku helper call. Public access was enabled after the trials. Agent-reported local checks and independent deployed checks are distinguished below. This single comparison does not establish a quality winner. The prompt did not require a particular sync transport. Update latency depends on polling phase and network conditions. Some emoji render as boxes in the screenshot environment. New Fable entries were added on 10 November 2026 using explicit Claude Code CLI runs and current frozen Neo skills. Their databases use the current Singapore development default; older entries use US databases. Public access was enabled by the host after completion. Fable 5.1 required same-session recovery after a missing test-harness keepalive. Its recorded time and token cost include recovery; this was not an uninterrupted trial.
Time taken
Execution time, seconds
Cost
Priced token subtotal, USD · * incomplete · — unknown
| Model | Harness | Run | Total | Input tokens | Output tokens | Cache read | Cache write | Reasoning | Original preview | Publication check | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-6.1 Sol | Codex | Complete | 585.9s | 597.0s | 80,716 | 20,373 | 2,629,565 | 0 | $0.6281 | medium | Protected during trial | Add + completion + reload passed |
| Claude Opus 5.5 | Claude Code | Complete | 742.8s | 752.9s | 6,313 | 33,858 | 6,190,105 | 128,675 | $2.5594 | medium | Protected during trial | Add + completion + reload passed |
| Claude Fable 5.2 | Claude Code | Complete | 744.8s | 761.4s | 82 | 38,293 | 2,682,055 | 103,195 | $3.1007 | medium | Protected during trial | Add + completion + reload passed |
| DeepSeek V4.1 Flash | DeepSeek Harness | Complete | 1276.4s | 1367.7s | 99,623 | 128,408 | 21,977,728 | 0 | $0.4735 | max | Protected during trial | Add + completion + reload passed |
| Claude Fable 5.1 | Claude Code | Recovered | 2069.7s | 4846.3s | 208 | 59,437 | 8,132,824 | 253,962 | $8.1817 | medium | Protected during trial | Live updates passed; mobile overflow |
Sorted by time taken, fastest first.
Claude Opus 5.5
742.8s · $2.5594
Add + completion + reload passed
Uses Neon Realtime. Original local browser checks used an origin bridge and two pages in one session. Publication checks used independent anonymous sessions against the real deployed app. Includes a Haiku helper call. Observed cross-session add/completion: 0.53s / 0.17s in one check, not a latency benchmark.
Custom skills supplied: 17
DeepSeek V4.1 Flash
1276.4s · $0.4735
Add + completion + reload passed
Uses Neon Realtime with a snapshot fallback. Original browser checks used polling; live-service checks ran in Node. No saved concurrent-add runtime check was found. Publication checks used independent anonymous sessions. Observed cross-session add/completion: 0.17s / 0.14s in one check, not a latency benchmark.
Custom skills supplied: 17
Claude Fable 5.1
2069.7s · $8.1817
Live updates passed; mobile overflow
Recovered in the same native session after the test setup omitted workspace keepalive. Execution and cost include the interrupted turn and recovery; total time also includes downtime. This was not an uninterrupted trial. Horizontal overflow at 390px viewport (504px document)
Custom skills supplied: 17