Claude Fable 5 nousi Googlen Android-paikannusmittarin kärkeen — Kustannukset, suorituskyky ja uusi arviointikehikko

Fable 5 also leads broader coding benchmarks industry-wide. Independent reporting shows it scoring 95.0% on SWE-bench Verified and 80.0% on SWE-bench Pro, well ahead of prior-generation models ^WLM. On the agentic coding benchmark Terminal-Bench 2.0, Fable 5 recorded an 84.3% accuracy ^WB. Multiple independent leaderboards now rank Fable 5 as the top overall AI model as of July 2026 ^BB.

Broader Standings: GPT 5.5, Gemini 3.1 Pro, and Others

All previously ranked models were reevaluated under the new Harbor-based methodology, producing a substantially refreshed leaderboard ^DA. The top ten according to Android Bench:

Source: 9to5Google's reporting from Android Bench's refreshed rankings ^A.

Key observations from the updated standings:

• Anthropic holds three of the top six spots (Fable 5, Sonnet 5, Opus 4.8), giving it the strongest overall lineup ^A.
• OpenAI's GPT 5.5 is the closest single-model challenger at 80.2, but trails by 4.3 points and has nearly double the latency of Fable 5 ^A.
• Google's own Gemini 3.1 Pro Preview ranks 5th at 73.7, behind both of Anthropic's top models and both OpenAI entries ^A. The lighter Gemini 3.5 Flash sits even lower at 71.1 with notably high cost and latency ^A.
• Kimi K2.7 Code offers the lowest cost ($48.10) but the lowest accuracy (70.4) among the top 10, illustrating the clearest cost-vs-accuracy tradeoff ^A.

Migration to the Harbor Framework

Google standardized Android Bench on the Harbor framework, an open-source evaluation ecosystem developed by the Laude Institute, the team behind Terminal-Bench ^DATG. Previously, Android Bench used a custom mini-swe-agent v1 harness. Harbor provides a standardized, container-based evaluation pipeline that supports cloud deployment, community task submission, and reinforcement learning rollouts ^ATQ.

The migration means all prior model scores are non-comparable — every score listed above is a fresh evaluation under the Harbor methodology ^9A. Google stated the move was necessary to keep evaluation standards state-of-the-art as LLMs rapidly improve ^D.

Introduction of Developer Contributions

For the first time, Google has opened Android Bench to community contributions ^9A. Developers can now:

• Submit real Android development tasks to be added to the benchmark dataset ^9A.
• Share their own benchmark evaluations of models using the same Harbor tooling, enabling direct comparison for individual use cases ^9A.

Google described this as a response to developer demand for "a way to provide feedback on our dataset" and a move toward deeper collaboration with the Android development community ⁹.

Cost-vs-Accuracy Tradeoffs

The refreshed leaderboard reveals a market with significant spread between cost leaders and accuracy leaders ^A:

• The highest-accuracy model (Claude Fable 5, 84.5) costs $133.20 per 1K tasks.
• The cheapest model in the top 10 (Kimi K2.7 Code, $48.10) scores only 70.4, a 14.1-point accuracy gap.
• Gemini 3.5 Flash is the most expensive option ($165.60) yet scores only 71.1.
• GPT 5.5 offers the best balance of high accuracy and moderate cost: 80.2 at $138.30.
• Claude Sonnet 5 provides a mid-range value play: 76.2 at $99.90.

Competitive Dynamics

The July 2026 update reinforces a three-way race among Anthropic, OpenAI, and Google ^A:

• Anthropic is the clear leader on pure accuracy. Fable 5's 84.5 score and Anthropic's depth with three models in the top 6 signal that its investment in Android-specific coding ability is paying off ^AM.
• OpenAI is the strongest single-model competitor with GPT 5.5 at 80.2, and GPT 5.4 at 74.1 holds a respectable position, but OpenAI lacks the lineup depth of Anthropic ^A.
• Google's Gemini line trails both competitors on accuracy. Despite owning the benchmark, Google's own models occupy 5th and 8th place — a competitive vulnerability the company is publicly acknowledging by opening the benchmark to the community and inviting third-party model submissions ^9AA.

This is the first Android Bench update where an Anthropic model leads Google's own benchmark for Android coding, a symbolic shift in the mobile AI assistant market.