Track record
TensorFeed has recorded 400 routing calls across four task types since August 8, 2026. Of the 40 calls re-checked 90 days later, 70% were still valid and 0% were still the pick. Every call that aged badly was a pick of Mistral Medium 3.5, which was later deprecated or removed from the catalog.
Every day at 07:00 UTC TensorFeed records the model its routing engine would pick for four task types, from a fixed question set that live traffic never touches. Each call is re-checked 7, 30 and 90 days later against what TensorFeed would pick that day and whether the original model is still supported.
How the calls held up
400 calls over 100 days, from August 8, 2026 to November 15, 2026. Superseded means a newer model now leads but the original pick is still available; it is not counted as a miss.
- Re-checked after 7 days100%still valid, of 372 calls
- Still the pick
- 53%
- Held
- 197
- Superseded
- 175
- Reversed
- 0
Under unchanged scoring constants (160 calls): 100% still valid, 62.5% still the pick.
- Re-checked after 30 days100%still valid, of 280 calls
- Still the pick
- 7.5%
- Held
- 21
- Superseded
- 259
- Reversed
- 0
Every call at this horizon was re-checked after at least one TensorFeed input refresh.
- Re-checked after 90 days70%still valid, of 40 calls
- Still the pick
- 0%
- Held
- 0
- Superseded
- 28
- Reversed
- 12
Every call at this horizon was re-checked after at least one TensorFeed input refresh.
Current picks
| Task | Pick today | Pick since | Distinct picks | Still valid at 90 days |
|---|---|---|---|---|
| Code | GPT-6.1 Sol | November 15, 2026 | 7 | 0% |
| Reasoning | GLM-5.3 | November 15, 2026 | 10 | 100% |
| Creative writing | GLM-5.3 | November 15, 2026 | 11 | 100% |
| General | GPT-6.1 Sol | November 15, 2026 | 10 | 80% |
Calls that aged badly
A call is reversed when the model it picked was later deprecated, scheduled for shutdown, or removed from the priced catalog. Repeated picks of the same model are grouped.
- Code: Mistral Medium 3.5
Picked from August 8, 2026 to August 17, 2026; 10 re-checks reversed. Reason: deprecated (sunset 2027-04-30). Every one of these re-checks also ran after a TensorFeed input refresh.
- General: Mistral Medium 3.5
Picked from August 16, 2026 to August 17, 2026; 2 re-checks reversed. Reason: deprecated (sunset 2027-04-30). Every one of these re-checks also ran after a TensorFeed input refresh.
Pick history by task
Code (11 runs)
- Mistral Medium 3.52026-08-08 to 2026-08-26 (19 days)
- Claude Sonnet 52026-08-27 to 2026-09-05 (10 days)
- GLM-5.32026-09-06 to 2026-09-14 (9 days)
- DeepSeek V4.1 Flash2026-09-15 to 2026-10-19 (35 days)
- Claude Sonnet 5.52026-10-20 to 2026-10-21 (2 days)
- GLM-5.32026-10-22 to 2026-10-22 (1 day)
- Claude Sonnet 5.52026-10-23 to 2026-10-29 (7 days)
- Claude Opus 5.52026-10-30 to 2026-11-12 (14 days)
- GPT-6.1 Sol2026-11-13 to 2026-11-13 (1 day)
- Claude Opus 5.52026-11-14 to 2026-11-14 (1 day)
- GPT-6.1 Sol2026-11-15 to 2026-11-15 (1 day)
Reasoning (17 runs)
- Claude Opus 52026-08-08 to 2026-08-15 (8 days)
- Gemini 2.5 Pro2026-08-16 to 2026-08-16 (1 day)
- Claude Opus 4.72026-08-17 to 2026-08-17 (1 day)
- Gemini 2.5 Pro2026-08-18 to 2026-08-21 (4 days)
- Claude Opus 4.72026-08-22 to 2026-08-22 (1 day)
- Gemini 2.5 Pro2026-08-23 to 2026-08-26 (4 days)
- GPT-5.6 Luna2026-08-27 to 2026-08-31 (5 days)
- LongCat-2.02026-09-01 to 2026-09-01 (1 day)
- GPT-5.6 Luna2026-09-02 to 2026-09-06 (5 days)
- Gemini 3.8 Flash2026-09-07 to 2026-10-19 (43 days)
- Claude Sonnet 5.52026-10-20 to 2026-10-21 (2 days)
- DeepSeek V4.1 Flash2026-10-22 to 2026-10-22 (1 day)
- Claude Sonnet 5.52026-10-23 to 2026-10-29 (7 days)
- Claude Opus 5.52026-10-30 to 2026-11-12 (14 days)
- GLM-5.32026-11-13 to 2026-11-13 (1 day)
- Claude Opus 5.52026-11-14 to 2026-11-14 (1 day)
- GLM-5.32026-11-15 to 2026-11-15 (1 day)
Creative writing (34 runs)
- Claude Sonnet 4.62026-08-08 to 2026-08-08 (1 day)
- Gemini 2.5 Pro2026-08-09 to 2026-08-12 (4 days)
- Claude Sonnet 4.62026-08-13 to 2026-08-13 (1 day)
- DeepSeek V4 Pro2026-08-14 to 2026-08-14 (1 day)
- Mistral Large2026-08-15 to 2026-08-15 (1 day)
- Gemini 2.5 Pro2026-08-16 to 2026-08-16 (1 day)
- Claude Sonnet 4.62026-08-17 to 2026-08-17 (1 day)
- Gemini 2.5 Pro2026-08-18 to 2026-08-21 (4 days)
- Claude Sonnet 4.62026-08-22 to 2026-08-22 (1 day)
- Gemini 2.5 Pro2026-08-23 to 2026-08-26 (4 days)
- DeepSeek V4 Pro2026-08-27 to 2026-08-30 (4 days)
- Gemini 2.5 Pro2026-08-31 to 2026-08-31 (1 day)
- DeepSeek V4 Pro2026-09-01 to 2026-09-01 (1 day)
- Gemini 2.5 Pro2026-09-02 to 2026-09-02 (1 day)
- DeepSeek V4 Pro2026-09-03 to 2026-09-04 (2 days)
- Gemini 2.5 Pro2026-09-05 to 2026-09-05 (1 day)
- Claude Opus 4.72026-09-06 to 2026-09-12 (7 days)
- Gemini 2.5 Pro2026-09-13 to 2026-09-14 (2 days)
- GLM-5.3 Flash2026-09-15 to 2026-09-17 (3 days)
- Gemini 3.8 Flash2026-09-18 to 2026-09-19 (2 days)
- GLM-5.3 Flash2026-09-20 to 2026-09-20 (1 day)
- Gemini 3.8 Flash2026-09-21 to 2026-09-21 (1 day)
- GLM-5.3 Flash2026-09-22 to 2026-09-22 (1 day)
- Gemini 3.8 Flash2026-09-23 to 2026-09-23 (1 day)
- GLM-5.3 Flash2026-09-24 to 2026-10-04 (11 days)
- Gemini 3.8 Flash2026-10-05 to 2026-10-05 (1 day)
- GLM-5.3 Flash2026-10-06 to 2026-10-19 (14 days)
- Claude Sonnet 5.52026-10-20 to 2026-10-21 (2 days)
- DeepSeek V4.1 Flash2026-10-22 to 2026-10-22 (1 day)
- Claude Sonnet 5.52026-10-23 to 2026-10-29 (7 days)
- Claude Opus 5.52026-10-30 to 2026-11-12 (14 days)
- GLM-5.32026-11-13 to 2026-11-13 (1 day)
- Claude Opus 5.52026-11-14 to 2026-11-14 (1 day)
- GLM-5.32026-11-15 to 2026-11-15 (1 day)
General (15 runs)
- Claude Opus 52026-08-08 to 2026-08-15 (8 days)
- Mistral Medium 3.52026-08-16 to 2026-08-23 (8 days)
- Gemini 2.5 Pro2026-08-24 to 2026-08-24 (1 day)
- Mistral Medium 3.52026-08-25 to 2026-08-26 (2 days)
- LongCat-2.02026-08-27 to 2026-09-05 (10 days)
- GLM-5.32026-09-06 to 2026-09-06 (1 day)
- Gemini 3.8 Flash2026-09-07 to 2026-09-28 (22 days)
- Gemini 3.1 Flash-Lite2026-09-29 to 2026-10-19 (21 days)
- Claude Sonnet 5.52026-10-20 to 2026-10-21 (2 days)
- GLM-5.32026-10-22 to 2026-10-22 (1 day)
- Claude Sonnet 5.52026-10-23 to 2026-10-29 (7 days)
- Claude Opus 5.52026-10-30 to 2026-11-12 (14 days)
- GPT-6.1 Sol2026-11-13 to 2026-11-13 (1 day)
- Claude Opus 5.52026-11-14 to 2026-11-14 (1 day)
- GPT-6.1 Sol2026-11-15 to 2026-11-15 (1 day)
How it is scored
- held
- The same model is still the pick for that task when the call is re-checked.
- superseded
- A different model leads today, but the original pick is still available and supported. The field moved; the call was not wrong when it was made.
- reversed
- The original pick has since been deprecated, scheduled for shutdown, or removed from the priced catalog. This is the label that means a call aged badly.
- unresolved
- The call could not be re-checked (no answer that day, or the question left the panel). Never counted in a rate.
- still_valid_pct
- held plus superseded, divided by held plus superseded plus reversed.
- still_the_pick_pct
- held, divided by held plus superseded plus reversed.
- same_methodology
- The same rates restricted to calls re-checked under unchanged scoring constants and an unchanged question set.
Each record carries a digest of the scoring constants it was computed under: the OpenRouter usage snapshot, the benchmark registry, and the task weights. TensorFeed refreshes the usage snapshot and the registry on a weekly cycle, so most calls re-checked 30 or 90 days later were re-checked under at least one refresh. Those outcomes are kept in the main rates and also broken out, so a change TensorFeed made to its own inputs is never presented as the market proving a call right or wrong. On November 15, 2026 the routing engine began settling contested near-ties (candidates within 5 percent of the top score) without the day-to-day latency term, after this record showed near-tie tasks changing pick almost daily; calls re-checked after that date carry the methodology change flag.
Get the data
The full record is free at /api/track-record. Add ?task=code (or reasoning, creative, general) for one task's day-by-day calls and labels. The routing engine itself is route-verdict, and the overall model scores it draws on are on the TensorFeed Intelligence Index.