AI model benchmarking platform Artificial Analysis has rolled out reasoning-effort filtering across its site. Users can now filter charts by reasoning level, toggle between a model's highest effort setting and its API default on the homepage, and compare scores across effort levels in the Model Release Comparison tool.
Reasoning Effort
推理努力度
A setting controlling how much thinking budget a model spends before answering, directly affecting scores and token usage.
Curated Intelligence
Alibaba launched Qwen4 Max, which scores 43 on the Artificial Analysis Intelligence Index at xhigh reasoning effort, a 9-point gain over Qwen3.8 Max and just 3 points behind Kimi K3.1 Pro. Cost per task drops to $3.27, roughly one-third of its predecessor and about 11% below Kimi K3.1 Pro, driven by 25% lower per-token pricing, 70% cheaper cache hits and 39% fewer output tokens.
AWS's official blog notes that even after prompt and application tuning, four request-level settings—cache duration, service tier, reasoning effort, and maxTokens—still materially change Amazon Bedrock's token volume and billing rates. It documents how each setting was tested and how to verify impact via response metadata, logs, and billing data.
Anthropic released Fable 5.2 and Mythos 5.2 today, with the llm-anthropic CLI plugin adding immediate support. Hands-on testing shows five reasoning levels from low to max, where xhigh produced the best SVG output while max exhausted 128,000 tokens on reasoning and failed to return an image, reproducing a known bug.
Artificial Analysis compared how many Intelligence Index tasks a fixed $100 budget buys: roughly 40 tasks with Opus 5.5 at high reasoning effort versus about 166 for DeepSeek V4.1 Flash and 2,500 for GPT-6 Luna, exposing a wide spread in per-task inference cost across frontier models. In parallel, Apple cut engineers' daily AI usage budgets from $300 to $200, with plans to drop to $100 and $25 for some teams, signalling that enterprises now treat cost-per-task as a core deployment metric. The figures are observational and come without disclosed methodology; no model release or capability upgrade is involved.
Andrej Karpathy used Claude Opus 5.5 (Medium effort) to scan 43 million Hacker News comments over two hours, filtering out 6,454 frequently recommended videos and launching a website called HackerTube. This case demonstrates the practical capabilities of frontier LLMs in autonomous data cleaning, long-context processing, and rapid full-stack application development.
OpenAI employee Ted Sanders explained the naming logic and performance differences in the GPT-6.1 series (Terra/Astra/Sol/Luna). He noted that while model size is a primary variable, newer architectures allow smaller models to outperform older larger ones. He provided simple selection guidelines based on price sensitivity: choose Astra for quality, Sol for balance, and Terra or Luna for extreme cost sensitivity.
Anthropic's Claude Sonnet 5.5 leads the Artificial Analysis Coding Agent Index, with its max effort mode scoring 68 points. However, the xhigh mode offers a compelling cost-performance trade-off, achieving 63 points at less than a quarter of the cost ($3.33 vs $14.19). The max mode boosts performance by generating 4.8× more output tokens and taking 3.4× more turns, yielding significant gains on benchmarks like Terminal-Bench 4.0.
SpaceXAI's Grok 4.7 is now available on Amazon Bedrock, emphasizing long-horizon task execution and self-verification. This post evaluates a tool-calling agent across three complex search-based tasks (outage diagnosis, scheduling, migration) by varying the 'reasoning effort' setting. The results highlight performance trade-offs at different deliberation levels and offer a practical framework for designing rigorous evaluations for agentic workloads.
A user leveraged the GPT-6 Astra agent to bypass Amazon Kindle's security restrictions, successfully installing a custom photo slideshow app. Previous attempts with GPT-5.6 Sol failed due to code errors, while Anthropic's Fable 5.1 and Opus 5.5 were blocked by built-in cybersecurity guardrails. This case highlights the varying capabilities and safety enforcement of different LLMs when tasked with modifying device firmware.