Notes from the arena
What the head-to-head runs actually show, how the methodology works, and takes you can disagree with by voting.
Release reports
GPT-5.6 Is Here: Sol, Terra and Luna Explained
· 5 minOpenAI’s GPT-5.6 ships in three tiers — Sol, Terra and Luna. What each is for, the pricing ladder, and how they perform in our live coding arena.
Claude Mythos vs Claude Fable 5: What’s the Difference?
· 6 minClaude Fable 5 and Claude Mythos 5 are the same Claude 5 flagship in two forms: one generally available with dual-use safeguards, one with safeguards lifted for approved organizations. Here is how access, safeguards and pricing differ — and where to watch Fable 5 run against Opus live.
Fable 5 is back — and the arena receipts are worth reading
· 6 minAnthropic’s Mythos-class model is usable again after the launch crunch. What 18 one-shot coding challenges, 3 compiled Godot games, and the community votes actually say about it.
Method notes
Takes & analysis
GPT-5.6, Grok 4.5, Muse Spark: Three Efficiency Claims This Week, Zero Common Yardstick
· 6 minOpenAI, xAI, and Meta each shipped a model this week with a headline efficiency number attached. None of the three numbers were measured the same way, which is exactly why blind arena votes matter more than vendor math.
The Week "Aggressive Pricing" Stopped Meaning Anything
· 5 minMeta, xAI, and OpenAI all shipped new models in a 48-hour span and all three called their pricing the headline. The actual per-token numbers don't agree with each other or with the adjectives.
Grok 4.5 Shipped With No Benchmarks. GPT-5.6 Shipped With Gamed Ones.
· 7 minTwo flagship launches landed one day apart this week, and neither gave you a number you could trust from the vendor. That is not a coincidence, it is the state of the industry.
Two Ways to Lose Trust: Models You Can't Test, Leaderboards That Grade Their Investors
· 6 minThis week's news cycle has Mythos and GPT-5.6 locked behind government clearance while Arena's leaderboard takes money from the labs it ranks. Both problems point at the same gap: independent, blind, live testing.
GPT-5.6 vs Claude Fable 5: Which Wins on Coding?
· 5 minGPT-5.6 vs Claude Fable 5: pricing, thinking-effort levels, token appetite and real cost-per-task — plus where to watch both flagships build from one prompt, live.
GPT-5.6 vs GPT-5.5: Is the Upgrade Worth It?
· 4 minGPT-5.6 vs GPT-5.5: what the Sol, Terra and Luna tiers change, how $5/$30 Sol compares to the $1.25/$10 flagship it replaces, and whether it one-shots better code.
GPT-5.6 vs Grok 4.5: Coding Head-to-Head
· 4 minGPT-5.6 vs Grok 4.5: OpenAI’s $5/$30 flagship against xAI’s $2/$6 price shot with 500k context. Pricing, speed and live one-shot coding outputs, community-voted.
GPT-5.6 vs Gemini 3: Same Prompt, Real Outputs
· 4 minGPT-5.6 vs Gemini: OpenAI’s $30 reasoning flagship meets Google’s fast, cheap Gemini 3 Flash on identical one-shot coding prompts — live outputs, community votes.
GPT-5.6 vs DeepSeek V4: Closed vs Open Weights
· 4 minGPT-5.6 vs DeepSeek V4: OpenAI’s $5/$30 flagship against an open-weight model at ~$0.14/$0.28. What a ~35x price gap buys on one-shot coding, judged live.
Grok 4.5 vs GPT-5.6: The Newest Flagships, Compared
· 4 minGrok 4.5 vs GPT-5.6: xAI’s $2/$6 flagship with 500k context meets OpenAI’s $5/$30 Sol on identical one-shot coding prompts — live outputs, community votes decide.
Claude vs ChatGPT for Coding: The Honest Comparison
· 4 minClaude vs ChatGPT for coding, compared honestly: same one-shot prompts, live outputs, blind community votes. Quality vs cost — and which to pick day-to-day.
Gemini vs ChatGPT for Coding: Speed vs Reasoning
· 4 minGemini vs ChatGPT for coding: Gemini 3 Flash’s speed vs GPT-5.5’s paid reasoning, run on identical one-shot prompts and judged live by blind community votes.
DeepSeek vs ChatGPT for Coding: Real Outputs Compared
· 4 minDeepSeek vs ChatGPT for coding: DeepSeek V4 Flash and GPT-5.5 compared on price, open weights, and real one-shot builds you can watch live and judge yourself.
ChatGPT vs Grok for Coding: Which Ships Better Code?
· 4 minChatGPT vs Grok for coding in 2026: GPT-5.5 and Grok 4.5 compared on price, speed, and one-shot code quality, with live community-voted head-to-head results.
Gemini vs Claude for Coding: Fast vs Frontier
· 4 minGemini vs Claude for coding: Google’s near-free Gemini 3 Flash meets Anthropic’s $50-output flagship Claude Fable 5 on identical one-shot prompts, voted live.
DeepSeek vs Claude: Open Weights Against the Flagship
· 5 minDeepSeek vs Claude for coding: V4 Flash’s open-weight $0.14/$0.28 pricing against Fable 5’s $10/$50 flagship — and where to watch both build the same prompt, live.
Qwen vs DeepSeek: The Open-Weights Coding Duel
· 4 minQwen vs DeepSeek: Qwen3.5 27B and DeepSeek V4 Flash go head-to-head in a live, community-voted coding arena. Open weights, FP8 serving, self-host value.
GPT vs Gemini for Coding: Which Should You Pick?
· 4 minGPT vs Gemini for coding: we run GPT-5.5 and Gemini 3 Flash on identical one-shot prompts, live in the browser, and let blind community votes pick the winner.
Grok vs GPT-5 for Coding: Real-World Outputs
· 4 minGrok vs GPT-5 for coding, settled with live one-shot outputs and blind community votes: Grok 4.5’s $2/$6 pricing and 500k context against GPT-5.5’s $1.25/$10.
Qwen vs Claude: Can Open Weights Match the Flagship?
· 4 minQwen vs Claude for coding: Alibaba's open-weight Qwen3.5 27B takes on Claude Fable 5 in live, one-shot arena matchups. Can a free download match a $50/M flagship?
Kimi vs Claude for Coding: Agentic Open Weights Tested
· 6 minKimi vs Claude for coding: Moonshot’s open-weight K2.7 Code faces Anthropic’s Fable 5 flagship on identical one-shot prompts, judged live by blind community votes.
DeepSeek vs Gemini for Coding: Value or Speed?
· 4 minDeepSeek vs Gemini: open-weight DeepSeek V4 Flash ($0.14/$0.28) meets Google’s fast Gemini 3 Flash on identical one-shot coding prompts — live outputs, community votes.
Kimi vs ChatGPT for Coding: Open Agentic vs Closed
· 4 minKimi vs ChatGPT for coding: Moonshot’s open-weight agentic K2.7 Code takes on GPT-5.5’s closed API. Pricing, self-hosting and live one-shot builds, judged by votes.
Claude vs DeepSeek: What $50 per Million Tokens Buys
· 4 minClaude vs DeepSeek on price: Fable 5’s $50-per-million output vs DeepSeek V4 Flash at $0.28 — a ~180x gap, what a finished one-shot app costs, live in the arena.
Grok 4.5 vs Claude Opus 4.8: Flagship Coding Test
· 4 minGrok 4.5 vs Opus 4.8: xAI’s $2/$6 newcomer against Anthropic’s $25-output vote leader. Pricing, effort levels and live one-shot coding builds — community-voted.
GLM-5.2 vs DeepSeek V4: Open-Weights Coding Kings
· 4 minGLM vs DeepSeek: GLM-5.2’s 744B MoE meets DeepSeek V4 Flash in a live, community-voted coding arena. Both open weights, both FP8 — self-host vs API economics.
Llama 4 vs GPT for Coding: Open Weights Against OpenAI
· 4 minLlama vs GPT for coding: Llama 4 Maverick and its open MoE weights face OpenAI's closed GPT-5.5 in a live, community-voted one-shot arena. Specs and stakes.
MiniMax vs Qwen: Open-Weights Coding Compared
· 4 minMiniMax vs Qwen: MiniMax M2.7 and Qwen3.5 27B face off in a live, community-voted coding arena. Open weights, MoE vs dense, FP8 serving, and the self-host math.
The Best AI for Coding in 2026, Judged Live in the Arena
· 9 minWhich AI is best for coding? Instead of quoting benchmarks, we give every model the same one-shot prompt, run the real outputs live, and let blind votes decide. Here is what the arena currently says — plus the best free options and picks for Python, vibe coding, and coding agents.
GLM-5.2 "beats" Claude Code. Same week, a different benchmark says Opus wins by 14 points
· 7 minThree benchmark stories about GLM-5.2 broke the same week, and they don't agree with each other. The disagreement is the actual story: single leaderboard numbers move with the harness, not just the model.
Karp Is Right That Tokens Get Wasted. He's Fixing the Wrong Thing
· 6 minPalantir's CEO says enterprises are paying for tokens that create no value, and he has a point. But the fix isn't hiding the meter behind credits, it's watching what a token spend actually produces.
Does thinking effort actually matter? Same model, low vs max
· 5 minOpus 4.8 and Sonnet ship at up to six effort levels each. The arena lets you blind-compare a model against itself — and the output data says effort is not the dial you think it is.