TREND TRACKING · AI Software Archive
Model capability advances while task economics and cash returns need separate proof
Global Model Watch | Initial baseline | November 10 2026
Fable 5.2 adds independent capability evidence. Google retains a preference-board advantage and Kimi expands open-weight deployment options. There is no single universal winner or automatic investment conclusion.
Fable 5.2 enters independent evaluation
Task tests and preference voting differ
Benchmark spend is not production cost per success
01 · RESEARCH BRIEF
The one-minute brief
Three deep checks cover Anthropic Fable 5.2, Google Gemini 4 Carbon and Moonshot Kimi K3.1 Pro. This is a sourced initial baseline, not a measured weekly change. AA v5.0.1 and the November 9 Arena text board are kept separate.
Audio transcript
This initial Global Model Watch separates capability, task economics and cash returns. Different benchmarks produce different leaders. Open weights widen deployment choices but licenses and operating costs still matter.
Known facts and open questions
- Series
- Global Model Watch
- Scope
- Three deep checks
- Baseline
- November 10 2026
Capability
Match version, effort and harness
Production
Measure success, retries, time and review
Commercialization
Verify payment, retention, revenue and cash ownership
| Model and effort | AA external index | Benchmark USD per task | Minutes per task |
|---|---|---|---|
| Claude Fable 5.2 max | 65 | 26.54 | 46.4 |
| Claude Fable 5.2 high | 62 | 9.08 | 17.6 |
| Gemini 4 Carbon high | 60 | 7.76 | 26.5 |
| GPT-6 Astra max | 57 | 6.24 | 23.0 |
| GPT-6.1 Sol max | 55 | 1.27 | 19.7 |
| Kimi K3.1 Pro | 46 | 3.68 | 37.7 |
02 · FACTS → IMPACT → VIEW
Why does this change matter?
Initial baseline
- What is confirmed
- No reproducible series snapshot found
- Why it matters
- Independent benchmarks, vendor announcements and license checks
- ACIS view
- Establish a prospective comparison point without inventing historical deltas
03 · EVIDENCE & ANALYSIS
Evidence and analysis
01｜Capability is not task economics
Fable 5.2, released on November 9, is already represented in independent testing. Higher effort brings higher benchmark scores and higher task spend; this is not evidence of better production economics. Fable and Mythos share an underlying model with different safeguards and access controls. Vendor coding and science results remain vendor evaluations.
02｜Preference leadership is not universal leadership
The November 9 Arena text board reports Carbon High at 1526±10, preliminary, with rank range 1–3; Argon High is 1522±5, also range 1–3. These intervals overlap. Arena preference and AA task evaluation are different measurements. A matched Fable 5.2 Arena entry was not verified. Google’s October 28 announcement described a phased restricted rollout; universal access was not independently confirmed.
03｜Open weights and economic ownership
K3.1 Pro is post-trained from Kimi K3 with expanded reinforcement learning, rather than a newly disclosed pretraining run from scratch. The official model repository and license are accessible. Some model-as-a-service uses require a separate agreement. Local hardware, operations and review costs remain unknown.
04｜Availability does not establish cash returns
Fable has paid-product and API distribution, but version-specific retention, revenue, margins and free cash flow are not disclosed. Comparable production and financial attribution are also missing for Carbon and Kimi. ARR, recognized revenue, commitments and cash received remain separate.
05｜Compute and distribution implications
Inference: longer reasoning and model routing can increase compute for valuable tasks while reducing expensive-model usage on simpler work. Efficiency can stimulate usage, so it does not mechanically reduce total GPU demand. Value capture by clouds, private deployers and applications still depends on contracts, customer ownership and reliability.
05 · VALIDATION & RISKS
What to watch next
Independent verification
Matched-version Fable 5.2 reliability and tool-use tests
What would weaken the view: Harness or safeguard differences explain the lead
Task economics
Matched-quality success, retry and human-review measurements
What would weaken the view: Longer inference or retries offset lower token prices
Commercial and funding evidence
Paid retention, recognized revenue, inference margins and cash attribution
What would weaken the view: Subsidized growth or obligations outrun cash
What would change our view?
Limited initial coverage. Rankings are snapshots and differences need not be statistically significant. Parameter and server counts are not value scores.
06 · FAQ
Key questions
Why no universal winner?
Tasks, versions, effort settings and confidence intervals differ.
Are open weights unrestricted?
No. Check the exact license and deployment costs.
Is benchmark task spend production cost?
No. Retries, tools, review and service costs remain additional.
07 · TERMS & SOURCES
Terms, sources and related research
Key terms
- AA
- Artificial Analysis independent evaluations
- Production cost
- Includes retries, tools and review
Public facts only; no personal portfolio or internal scoring. Necessary excerpts only, not a replicated leaderboard. Missing evaluation dates, effort labels and intervals remain unknown.