Between June 30 and September 3, 2026, Anthropic shipped Claude Sonnet 5, Opus 5 and Fable 5.1, and OpenAI shipped the GPT-5.6 family and GPT-6 Astra. The version of this page published in April was badly out of date, so this is a full rewrite. Neither product is objectively better. They have converged more than most people realize, and the differences left are specific enough to test. Specs come from each vendor's docs; every benchmark is labeled with who reported it.

The Lineups in September 2026

Anthropic sells four tiers. Fable 5.1 (September 1) is the top model, built for hard reasoning and long agent runs. Opus 5 (July 24) is the one Anthropic's docs tell you to start with. Sonnet 5 (June 30) is the fast middle tier, and its $2/$10 launch price is now permanent. Haiku 4.5 is the budget pick. Mythos 5.1, the same model as Fable 5.1 with different safeguards, is limited to vetted security and life-science teams (background in our Project Glasswing piece).

OpenAI's flagship is GPT-6 Astra (September 3), released first to its Daybreak cybersecurity program, then to ChatGPT Plus, Pro, Business and Enterprise and the API. Below it sits July's GPT-5.6 family: Sol, Terra and Luna. Luna has been the default for ChatGPT's free users since August.

Aspect Claude ChatGPT / OpenAI
Top model (API, per MTok in / out) Fable 5.1, $10 / $50 GPT-6 Astra, $10 / $50
Workhorse Opus 5, $5 / $25 GPT-5.6 Sol, $4 / $20 (promo through at least Nov 21)
Mid tier Sonnet 5, $2 / $10 GPT-5.6 Terra, $2 / $12
Budget tier Haiku 4.5, $1 / $5 GPT-5.6 Luna, $0.20 / $1.20
Context window 1M tokens, one flat rate (Haiku 4.5: 200K) 1.05M tokens, surcharge above 272K
Knowledge cutoff Jun 2026 (Fable 5.1), May 2026 (Opus 5) Apr 30, 2026 (Astra), Feb 16, 2026 (GPT-5.6)
Consumer plans Free; Pro $20; Max from $100 Free; Go $8; Plus $20; Pro $100 or $200
Coding agent Claude Code (Pro and up) Codex (Plus and up)

Context Windows Are No Longer Claude's Moat

This used to be the headline difference. Now Opus 5, Sonnet 5 and Fable 5.1 take 1M tokens, every current OpenAI model takes 1.05M, and both cap a response at 128K output tokens. The gap is in billing. Anthropic charges one per-token rate across the whole window. OpenAI bills any prompt over 272K input tokens at double the input rate and 1.5x the output rate for the entire request. If you routinely feed a monorepo or months of logs into one call, do that math first.

Token counts don't match across vendors either. Anthropic says the tokenizer it introduced with Opus 4.7 produces roughly 30% more tokens for the same text, which puts 1M tokens at about 555,000 English words. Compare cost per finished task, not the price sheet.

Coding Agents: Claude Code vs Codex

For most developers this is the comparison that matters now.

Claude Code runs in the terminal, in VS Code and JetBrains, in a desktop app and in the browser, with Slack and GitHub Actions integrations. It comes with Pro, Max, Team and Enterprise (not Free) or runs on an API key. It reads a CLAUDE.md for project instructions, and an AGENTS.md if your repo has one. The client is proprietary.

Codex offers a CLI, an IDE extension, cloud tasks and a desktop app for macOS and Windows, signed in through ChatGPT Plus, Pro, Business, Edu or Enterprise, or an API key. The Codex CLI is open source under Apache-2.0. OpenAI's $100 Pro tier, added in April, is aimed at heavy Codex users with 5x the Plus usage. With Astra, Codex also stopped leaning only on compaction for long sessions: the model keeps notes across context windows and can search earlier ones.

Day to day, the model and your plan's usage limits matter more than the harness. MCP, the tool standard Anthropic started, works in Claude Code and in OpenAI's API, and shared AGENTS.md support means one repo can serve both agents.

Benchmarks, With the Fine Print

Every number below is vendor-reported, on benchmarks the vendor chose. Anthropic's Fable 5.1 post predates Astra by two days, so most head-to-heads come from OpenAI.

Benchmark Claude OpenAI Reported by
Terminal-Bench 4.0 Fable 5.1: 55.8% GPT-6 Astra: 57.9% OpenAI (Anthropic reports the same 55.8%)
Terminal-Bench Science 0.1 Fable 5.1: 52.6% GPT-6 Astra: 64.6% OpenAI (Anthropic reports the same 52.6%)
Agents' Last Exam Opus 5: 55.5% GPT-6 Astra: 59.3% OpenAI
CursorBench 3.2 Fable 5.1: 73.4% GPT-5.6 Sol: 67.2% Anthropic (no Astra result)

Where both companies report the same benchmark, their Claude figures agree. The lead still depends on who is counting. OpenAI adds that Astra hit its Terminal-Bench 4.0 score at about 63% lower estimated cost per task than Fable 5.1, while Anthropic's restricted Mythos 5.1 posted 60.9% there, on a model most developers can't call. Use these as a reason to run your own evals.

API Experience

The APIs have converged. Anthropic's Messages API and OpenAI's Responses API both offer function calling, structured outputs, hosted web search, code execution, computer use, MCP tools, prompt caching and half-price batch jobs. Cached input costs a tenth of the normal rate on both, and Fable 5.1 cache reads drop to $0.25 per million tokens. Both sell a faster mode at double price. Distribution is no tiebreaker either: Claude runs on Anthropic's API, Amazon Bedrock, Google Cloud and Microsoft Foundry, and OpenAI's models run on its API, Azure and now Bedrock.

Safety, Refusals and Published Rules

Both companies publish their rules. Anthropic released a rewritten constitution for Claude in January 2026 under CC0, and OpenAI's public Model Spec was last revised in August. Either tells you more about refusals than a forum thread. Both also gate their strongest cyber capabilities: Mythos 5.1 goes only to vetted teams, and OpenAI calls Astra its first model to reach its internal "Critical" cybersecurity threshold. On the API, Opus 5 and Fable 5.1 can return a refusal stop reason, and Anthropic can retry on a fallback model; in Claude Code, requests Opus 5 flags fall back to Opus 4.8 by default. Treat refusals as a normal outcome on either platform, not a crash.

Web Search and Everyday Chat

The old gap on current information is gone: Claude's free plan searches the web, creates files and runs code, and ChatGPT's free tier searches too. Both let you trade speed for depth: ChatGPT added a thinking slider in August, and Claude runs at adjustable effort levels (Fable 5.1 defaults to medium on Claude.ai, high in Claude Code). ChatGPT also generates images; Claude's models output text. Writing style is taste, so try both on documents you actually write.

When Claude Fits Better

  • Very long prompts, where one flat rate beats a surcharge past 272K tokens
  • A capable mid-tier cheaply: Sonnet 5 at $2 in, $10 out
  • Repeated large context on the top model, with $0.25 cache reads
  • Teams already built around Claude Code's CLAUDE.md files, skills and hooks

When ChatGPT Fits Better

  • The strongest published coding and computer-use numbers today (Astra, vendor-reported)
  • The cheapest capable tier: Luna's $0.20 input is a fifth of Haiku 4.5's
  • An agent CLI you can read, audit and fork
  • Image generation next to the chat model, and more plan price points

The Practical Recommendation for Developers

Test before you commit. Run ten real tasks from your backlog through Opus 5 and GPT-5.6 Sol (or Fable 5.1 and Astra for the top tier) and record cost per finished task, retries and time. Both $20 plans include a coding agent, so a month of each is a cheap eval. Keep your setup portable with an AGENTS.md, MCP servers and prompts that don't depend on one model's quirks. Five major launches in just over nine weeks says the lead will change hands again.

Conclusion

As of late September 2026, OpenAI leads the head-to-heads it published with Astra, sells the cheapest capable model and ships an open-source agent CLI. Anthropic offers flat pricing across 1M tokens, an aggressively priced Sonnet 5 and cheap cache reads on its top model. Neither is universally better, and both will look different by December. Evaluate on your own work and skip the tribal debate.

Watch the AI Tools You Depend On

Track live status for ChatGPT, Claude and other AI agents, or check whether an API is down before you start debugging your own code.

AI Agent Tracker Service Status

About Signal

Signal is TerminalFeed's Data and AI Editor, covering AI agents, LLMs, and machine learning at scale. They test and evaluate AI tools for practical application and help teams understand the tradeoffs between competing solutions.