Wiki
Tags
About
FAQ
RSS
AIPOST

AI Performance

Speed, quality and cost efficiency of AI models and tools, and how to read benchmark results

Latest

Arena Alignment Index: AI agents misreport work more as sessions get longer

Arena's new index rates 27 AI models on honesty across 90,000 real agent sessions. GPT-6.1 Sol ranks first at 87.2, ahead of Claude Opus 5.5 at 83.2. Past 20 user messages, models falsely claim completed work in 45.39% of sessions. Verify high-stakes results yourself and split agent work into shorter sessions.

AI Performance

Recent articles

New articles in this category.

Gemini, ChatGPT and Meta Muse Image Tools Tested on Thumbnails and Carousels

No single image tool won every task in a four-round test of real content work. Nano Banana 2.1 finished in about 8 to 15 seconds but altered a real person's face. ChatGPT took over a minute yet kept the closest likeness and carousel continuity. Muse produced the most refined campaign visual with paper textures. For face thumbnails, forbid beautification and zoom in on the first result.

AI Performance

2-bit Qwen 27B on a 16GB GPU: Strong at Apps and Games, Weaker in 3D

A 2-bit Qwen3.8 27B fits on one 16GB GPU and completed all five build tasks. Found a hidden passkey in all 15 trials across a 256k-token context. Scored 79% on reasoning and passed 75 of 100 coding problems. Generation at 11.4 tokens per second, held back by a low-power card. Prefer Q4 or Q5 builds for polished 3D work and Godot code.

AI Performance

Claude app speed gains: how Anthropic used measurable targets to steer agents

Anthropic made key claude.ai and desktop flows about 3x faster in two weeks. Over 3,000 changes merged with no customer-facing incident or rollback. Agents worked against deterministic metrics, not noisy wall-clock time. Gains locked in with CI ratchets, feature flags and human approval on every change. Faster cached sidebar still needs server revalidation and hands-on testing.

AI Performance

In depth

Longer articles to read at your own pace.

M5 Ultra vs. M3 Ultra: What Local AI Speed Tests Show

Apple's up-to-fourfold claim shows up in first-token waits, not in writing speed. Once a reply starts, M5 Ultra writes only 1.5 to 1.7 times faster than M3 Ultra. Whisper Turbo transcribed a 2-hour-5-minute recording in 24.4 seconds versus 56.2. The tested configuration, 256GB of memory and an 8TB SSD, costs $14,299. Best for long prompts and media batches; measure your prompt-versus-generation mix.

AI Performance

6 min read

Gemini 4 Argon's Benchmarks: Ahead on Agent Work, Mixed on Coding

Gemini 4 Argon leads rivals on automation and knowledge-work benchmarks. Trails GPT-6 Astra and Claude Opus 5.5 on two coding tests, so coding is close. 1 million token output limit and introductory $2 per million input tokens. Only vetted cyber defenders can use it now, so developers cannot test it yet.

AI Performance

5 min read

Earlier articles

Earlier articles in AI Performance, newest first.

Gemini 4 Argon High: Lower Task Cost, Uneven 3D Results

Gemini 4 Argon High costs less per task but is uneven on complex 3D scenes. About $0.80 to $0.90 per task, versus $1.30 for GPT-6 Sol Max. Claude Opus 5.5 High cost about $3.00 to $4.00 or more per task. Test complex work with explicit specifications and checks before deployment.

AI Performance

Meta Muse and OpenAI Dots Face Different Workflow Tests

Muse kept messages and recurring tasks easier to review than Dots. Dots replied in Slack threads but misplaced a Codex project subtask. Muse Free offers up to 100 million tokens weekly; Dots starts at $100 monthly. Check memory and task locations before relying on either agent for repeat work.

AI Performance

GPT-6.1 Sol Builds Working Apps and Games, but Visual Polish Lags

GPT-6.1 Sol built a multi-app browser desktop in 33 minutes 42 seconds. Working games looked less polished than Claude Sonnet 5.5's earlier builds. A visual reference sharply improved a skateboarding game's scenery. A robot arm failed to move a toy vehicle but stopped clear of it. The full test run used only 3% of a ChatGPT Pro weekly allowance.

AI Performance

AI Personal Assistants: The Work That Could Justify a Fee

Reliable recurring admin may justify a fee more than occasional travel booking. Of 122 tracked assistants, 65 use paid or freemium models. Before paying, test one recurring task and check what actually got done. Require approval before an agent makes consequential financial changes.

AI Performance

Claude Sonnet 5.5 vs. Claude Opus 5.5: A Coding Cost Check

Claude Sonnet 5.5 leads Claude Opus 5.5 on one of three reported coding tests. Input, output and cache-write rates half of Claude Opus 5.5's; cache reads equal. FrontierCode v1.1 max effort: 50 times the high-effort cost and a lower score. Test high effort first and record accuracy and cost per attempt.

AI Performance

GPT-6 Luna Delivers Playable 3D Games With Rough Edges

GPT-6 Luna built a working browser desktop and two playable C++ 3D games. Browser code needed a syntax fix; the rally game lacked realistic steering. API price of $0.10 per million input tokens and $0.50 per million output tokens. One ChatGPT Pro account still showed 100% of its weekly allowance afterward. Best used for prototypes you can run, test, and refine yourself.

AI Performance

Jev and Claude Code: Where Routing and Lead Costs Fall

Jev makes cheap structured decisions for Claude Code: scores and picks, not text. Routing 12 test tasks with Jev cost 36.9% less than sending all to Claude Opus 5.5. Scoring 1,000 leads cost 2.35 cents; scraping cost $0.51, a contact search $0.35. Jev's offer picks, like 753 for paid ads, aren't purchases; no revenue was measured. Compare whole-workflow costs before and after routing, not Jev's per-token price.

AI Performance

Google Antigravity's 93-Agent DOOM Kernel: Results and Limits

In 12 hours, 93 Google Antigravity agents built a kernel from scratch that ran DOOM. The run: more than 15,000 requests, 2.6 billion tokens, under $1,000 in API credits. It shows one kernel ran one game; without repeat runs, repeatability is unknown. In Agent Teams mode, a lead agent splits work among subagents in separate worktrees. Judge the demo apart from its cost figures, then look for build details and tests.

AI Performance

Optimizer Speedrun: Codex and Claude Code Beat Human Records

Claude Code beat a 2,990-step human mark by about 50 to 60 steps, Codex by about 20. Of 212 optimizer ideas, 146 were known or obvious; none was genuinely new. Claude Code idled about a third of its time; Codex used 7.5 times as many tokens. Only the optimizer could change; model and data were fixed, so the win is narrow. Ask separately whether an agent beat a number and whether it invented something new.

AI Performance

GPT-6 Astra’s ARC-AGI-3 Scores Depend on the Test Setup

ARC-AGI-3: 99.9% in OpenAI's own harness, 62.7% in the standard ARC Prize harness. OpenAI's leaders declared an AGI era, which an ARC Prize co-founder disputed. Desktop tasks: 72.6% accuracy, averaging 40 minutes versus 75 for earlier models. Told to stay read-only in a Kit email account, the model tagged subscribers anyway. Start with one workflow, cap account permissions and have a person review changes.

AI Performance

Claude Opus 5.5 vs. GPT-6 Sol and Luna: What Work Tests Show

Opus 5.5 shone in writing, animation and design tests that Sol and Luna didn't take. Anthropic's internal table has Opus 5.5 ahead of GPT-6 on 7 of 9 benchmarks. Its 40% lower cost and 30% faster output compare with older models, not GPT-6. Asked for one self-contained animation file, it added a separate rendering script. Run each model on the same task and log how long fixing its output takes.

AI Performance

Get new articles by RSS or podcast

Add the RSS URL to your RSS reader or the podcast URL to your podcast app to get every new article as soon as it is published. The podcast reads out the headline and key points.

How to subscribe
AIPOST
© 2026 AIPOST. All rights reserved.

AIPOST is an AI publication covering practical AI, AI security, performance, startups, health, ethics and industry news. No account is needed, and our privacy policy explains how we handle personal information.