ModelsArchitectures & capability
Anthropic's Sonnet 5.5 nears frontier scores at one-fifth of Fable 5.1's per-token price, but uses far more tokens per task
Anthropic's Sonnet 5.5 costs $2/$10 per million tokens and scores two points behind Opus 5.5 on an independent index. Heavy token use at high effort reduces the per-token savings, and the claim that frontier models will soon run on laptops lacks support.
AT Pulse research deskOctober 6, 2026 · 5 min read · AI-assisted analysisIn four weeks, Anthropic released three models at falling prices. Fable 5.1 launched at $10/$50 per million input/output tokens. Opus 5.5 followed at $4/$20, and Sonnet 5.5 at $2/$10, which is one-fifth of Fable's per-token price. Artificial Analysis independently ranks Sonnet 5.5 second of 225 models on its Intelligence Index, two points behind Opus 5.5 at max effort. [2] [6] [7] [4]
Sonnet 5.5's lower token price does not guarantee a lower cost per task. Across its five effort levels, its index score rises from 36 to 56, while output tokens rise from about 23M to 420M. At max effort it uses about 60% more tokens than Opus 5.5 and roughly seven times as many as GPT-6 Astra per task. [10] [11] [7]
Anthropic reports that Opus 5.5 beats Fable 5.1 on Terminal-Bench 4.0 and CursorBench 4.0. These results come from Anthropic itself. No public comparison runs both models on the same test setup, and Anthropic chose which customer results to cite. [3]
The commentary video draws two inferences that the sources do not support. It argues that cheaper models must be smaller and that frontier AI will run on laptops, but Anthropic has not disclosed parameter counts. Its AVBD physics demo also proves less than it suggests. Sonnet 5.5 ported partial, already-public source code into a single HTML file rather than rebuilding the method from scratch. The prompts, number of attempts and success criteria were not published. [1] [10] [5]
independent scoring shows near-frontier capability at one-fifth of Fable 5.1's per-token price. Implication: teams should budget per task and tune effort levels, since token use varies about 18-fold.
Read the full assessment
Falling API prices do not mean frontier models will run locally or with open weights.
Executive brief
Within four weeks, Anthropic cut the per-token price of frontier performance by about 80%. Claude Fable 5.1 launched on 1 September 2026 at $10/$50 per million input/output tokens (MacRumors, VentureBeat). Sonnet 5.5 followed on 28 September at $2/$10. Independent testers place it two points behind Opus 5.5 (Artificial Analysis). Two Minute Papers takes this as evidence that frontier AI will shrink to laptops (video, 1:43). That inference rests on undisclosed model sizes. Sonnet 5.5 also uses a record number of tokens per task, which reduces its per-token savings.
What changed and event timeline
AVBD published
Giles, Diaz and Yuksel present Augmented Vertex Block Descent at SIGGRAPH 2025. It is a GPU physics solver that handles millions of colliding, jointed objects in real time. The 2D and 3D demo code is public.
Fable 5.1 launches
It is the same model as the restricted Mythos 5.1 but with production safeguards. Anthropic reports it beats Fable 5, Opus 5 and GPT‑5.6 Sol, and that cache-read cuts make it 25–45% cheaper than Fable 5.
Opus 5.5 at 40% of Fable's price
Priced at $4/$20, Anthropic says it matches Fable 5.1 on most work and calls it the strongest model it has tested.
Sonnet 5.5 ships
Released as
claude-sonnet-5-5at an unchanged $2/$10. Anthropic says it runs more than 30% faster than Sonnet 5.More detail
Commentary video
Two Minute Papers demos Sonnet 5.5 porting AVBD to a single HTML file. It argues that intelligence depends on training quality rather than parameter count.
Capabilities and access
- Sonnet 5.5 (
claude-sonnet-5-5): $2/$10. Available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry (search summary of cellcog/MarkTechPost coverage). - Opus 5.5: $4/$20. It carries Fable-level safeguards that restrict exploit development and bioweapons misuse (TechCrunch).
- Fable 5.1: $10/$50, generally available. Mythos 5.1 is the same model, limited to trusted-access programs (MacRumors).
Read the full section
- Sonnet 5.5 (
claude-sonnet-5-5): $2/$10. Available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry (search summary of cellcog/MarkTechPost coverage). It has five effort levels, from low to max (Artificial Analysis). - Opus 5.5: $4/$20. It carries Fable-level safeguards that restrict exploit development and bioweapons misuse (TechCrunch).
- Fable 5.1: $10/$50, generally available. Mythos 5.1 is the same model, limited to trusted-access programs (MacRumors).
Technical analysis for researchers and developers
- On the Artificial Analysis Intelligence Index, Sonnet 5.5 scores 36 at low effort, 41 at medium, 47 at high, 52 at xhigh and 56 at max.
- Anthropic has not disclosed parameter counts (Artificial Analysis). The video's claim that cheaper means smaller is its own inference (1:46).
- The video says Sonnet 5.5 ported partial public code, not the whole method (0:37).
Read the full section
- Effort scaling. On the Artificial Analysis Intelligence Index, Sonnet 5.5 scores 36 at low effort, 41 at medium, 47 at high, 52 at xhigh and 56 at max. Output tokens rise from 23M to 420M across those settings; the median model uses 81M (Artificial Analysis). At max effort it uses about 193k output tokens per task, roughly 7× GPT‑6 Astra (Artificial Analysis on X).
- Model size. Anthropic has not disclosed parameter counts (Artificial Analysis). The video's claim that cheaper means smaller is its own inference (1:46).
- AVBD demo. The video says Sonnet 5.5 ported partial public code, not the whole method (0:37). The public demos already exist (Utah), so this tests porting more than reproducing the paper from scratch. The prompts, number of attempts and success criteria are not published.
Claims and evidence
- Opus 5.5 beats Fable 5.1 on Terminal-Bench 4.0 (66.4% vs 55.8%) and CursorBench 4.0 (57.8% vs 51.8%) () — Vendor-reported
- Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, ahead of Opus 5.5's 66.4% () — Vendor-reported
- Sonnet 5.5 is 2 of 225 on the Intelligence Index, 2 points behind Opus 5.5 at max effort () — Independent
Read the full section
| Claim | Status |
| Opus 5.5 beats Fable 5.1 on Terminal-Bench 4.0 (66.4% vs 55.8%) and CursorBench 4.0 (57.8% vs 51.8%) (VentureBeat) | Vendor-reported |
| Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, ahead of Opus 5.5's 66.4% (Decrypt) | Vendor-reported |
| Sonnet 5.5 is #2 of 225 on the Intelligence Index, 2 points behind Opus 5.5 at max effort (Artificial Analysis) | Independent |
| Sonnet 5.5 and Opus 5.5 effectively tie on GDPval-AA (1844 vs 1846) (Decrypt) | Independent (Artificial Analysis) |
| Frontier models will run on laptops (video, 2:26) | Speculation; no corroboration found |
Context and prior work
- AVBD extends Vertex Block Descent with an augmented Lagrangian formulation.
- Developers prompted earlier Claude models to implement AVBD in 2025 (Renaud on X).
- Sonnet 5.5 replaces Sonnet 5, released in June 2026 (Decrypt).
Read the full section
- AVBD extends Vertex Block Descent with an augmented Lagrangian formulation. This lets it handle hard constraints with infinite stiffness while staying stable at low iteration counts (Utah, 80.lv).
- Developers prompted earlier Claude models to implement AVBD in 2025 (Renaud on X).
- Sonnet 5.5 replaces Sonnet 5, released in June 2026 (Decrypt).
Limitations, safety and contested findings
- Token use offsets the price cut.
- Weak independent evidence on Opus vs Fable.
- Anthropic reports Opus 5.5 makes 85% fewer attempts to get around its boundaries, but acknowledges that reliably catching every failure before deployment is unsolved (same source).
Read the full section
- Token use offsets the price cut. At max effort, Sonnet 5.5 uses about 60% more tokens than Opus 5.5, which narrows its per-task cost advantage (Decrypt).
- Weak independent evidence on Opus vs Fable. VentureBeat notes there is no public same-harness comparison, and that the customer results Anthropic cited were selected by Anthropic (VentureBeat).
- Safety. Anthropic reports Opus 5.5 makes 85% fewer attempts to get around its boundaries, but acknowledges that reliably catching every failure before deployment is unsolved (same source). METR ran pre-release testing; its findings were not detailed in the coverage (TechCrunch).
- Sponsorship. The video carries a sponsored segment for Lambda GPUs (3:07).
Business and practitioner implications
- Sonnet 5.5 at medium or high effort may give the best cost per task. Max effort can cost as much as Opus.
- Budget per task, not per token. Token use varies about 18× across effort levels (Artificial Analysis).
- Don't plan around running frontier models locally. These models remain API-only and proprietary, and their sizes are undisclosed.
Read the full section
- Route by tier, then tune effort. Sonnet 5.5 at medium or high effort may give the best cost per task. Max effort can cost as much as Opus.
- Budget per task, not per token. Token use varies about 18× across effort levels (Artificial Analysis).
- Don't plan around running frontier models locally. These models remain API-only and proprietary, and their sizes are undisclosed. The "advantage disappearing" thesis currently means falling API prices, not open weights.
Sources
Read the full section
- Two Minute Papers video
- TechCrunch: Opus 5.5
- VentureBeat: Opus 5.5 vs Fable 5.1
- MacRumors: Fable 5.1
- Decrypt: Sonnet 5.5
- Benzinga: Sonnet 5.5
- cellcog: Sonnet 5.5 release
- Artificial Analysis: Sonnet 5.5
- Artificial Analysis on X (score)
- Artificial Analysis on X (tokens)
- Utah AVBD project page
- 80.lv: AVBD
- Renaud on X
The source trail.
Sources (13)
The Billion Dollar AI Advantage Is Disappearing
Transcript retrieved via youtube_auto_captions; language en. Timestamped text, not direct audiovisual review. Source: https://www.youtube.com/watch?v=ZHVNTTKu9fU. Automatic captions/transcription may contain errors.
www.youtube.com