One-Sentence Answer
Earlier this week -- on 22 September 2026 -- OpenAI cut GPT-6 Sol and Luna API prices by 50% and Anthropic released Claude Opus 5.5 at 40% lower running cost, within hours of each other. Frontier-class intelligence now costs roughly half what it did the day before, and the number that decides your stack is cost per completed task, not who sits at the top of the leaderboard.
This is one of those weeks where the ground shifts quietly. I watched a daily AI news roundup covering the same launches (the source video is linked in the takeaways below), then went to the primary sources -- OpenAI's and Anthropic's own launch posts -- because the headline numbers are worth checking line by line.
The Problem: "Just Use the Best Model" Stopped Being Good Advice
For most of 2026 the advice was simple: pick the smartest model you can afford. Two things have broken that:
- The smartest model changes weekly. GPT-6 Astra, Claude Fable 5.1, Mythos 5.1, Opus 5, Grok 4.7 -- the top of the table rotates faster than you can retrain your habits.
- Agentic work multiplies every token you spend. A chat prompt costs what it costs. An agent that plans, calls tools, reads a file, reflects and revises can replay the same 100,000-token context a dozen times per job. Your bill is not the prompt -- it is the prompt times the number of steps.
Both vendors have noticed. Neither launch this week was about being smarter. Both were about being cheaper per finished job.
What Actually Shipped
1. GPT-6 Sol and GPT-6 Luna (OpenAI)
GPT-6 Astra arrived earlier this month as OpenAI's flagship. Sol and Luna bring that generation's training methods down the price curve:
- GPT-6 Sol: $2 per million input tokens, $10 per million output tokens (was $4 / $20) -- 50% cheaper in both directions.
- GPT-6 Luna: $0.10 in, $0.50 out (was $0.20 / $1.20) -- a 95% discount versus Claude Sonnet 5's $2 / $10 rates.
OpenAI confirmed to VentureBeat that these are permanent prices, not promotional. Note where Sol lands: exactly on Claude Sonnet 5's rate, at half of Opus 5.5's.
The performance claims use cost per task, not raw benchmark rank:
- On AutomationBench (business workflows across 47 tools), GPT-6 Sol at xhigh effort scores 33.2% at $0.27 per task, beating Claude Opus 5 at max effort, which costs 11.1x more per task.
- On Agents' Last Exam it scores 56.4% at max effort, above Claude Opus 5's best, at 60% lower cost per task.
- On DeepSWE v1.1 (real codebase engineering), Sol max reaches 68.8% -- within 1.1 points of Claude Fable 5's best (69.9%) at roughly 80% lower cost per task. Luna max scores 66.6%, 93% cheaper per task than Opus 5.
- On OSWorld 2.0 offline (computer use), Sol xhigh matches Opus 5 medium, 60.5% vs 60.3%, at about 80% lower cost per task.
Two practical details matter more than the charts. First, OpenAI improved prompt caching for GPT-6: cached input reads get a 90% discount, and reasoning effort or tool availability can now change mid-conversation without invalidating the cache. Second, availability is staged: both models are in ChatGPT Work and Codex for paid plans, Luna is available to Free and Go users in the desktop app, and both are live in the API as gpt-6-sol and gpt-6-luna. They are not yet in Chat.
2. Claude Opus 5.5 (Anthropic)
Same day, Anthropic released the first model of its new Claude 5.5 family:
- $4 per million input tokens, $20 per million output. Cache reads drop from $0.50 to $0.20 per million -- 60% cheaper than Opus 5.
- Anthropic says that at default settings it costs 40% less than Opus 5 on typical workloads, because it uses fewer tokens to finish the same work, not just a lower sticker price. Output generation is more than 30% faster.
- Benchmarks: Terminal-Bench 4.0 at 66.4% (Fable 5.1: 55.8%, GPT-6 Astra: 57.9%), FrontierCode v1.1 Main at 54.4%, and 1846 Elo on GDPval-AA v2.1, a benchmark spanning 44 occupations.
- On FrontierCode at default effort it beats GPT-6 Astra's top score for roughly 20% of the cost per task; on CursorBench it beats GPT-5.6 Sol by 11 points at about a third of the cost.
- Anthropic also confirmed that Claude Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, with the same efficiency and safety work, and raised subscription usage limits while letting users bank a rate-limit reset for whenever they want it.
3. Three More Signals From the Same Week
- Alibaba teased the Qwen 4 family at the Apsara conference: Qwen 4 Max, Plus, Flash, and a Qwen 4 27B model that matters because it is the size class you can run locally. The models are still in training.
- Tencent's Hy Image 3.5 preview went live at roughly $0.024 per image (0.15 RMB for 2K on the mainland API), with reference images and failed generations not billed -- image generation is following the same cost curve as text.
- The naming fight. At the UN on 22 September, President Trump directed US government documents to refer to AI as "super intelligence" (SI), arguing the word "artificial" makes the technology sound fake. It is a language story rather than a technical one, but if you write or sell in this space, the vocabulary is about to move.
The Arithmetic That Matters: Cost per Task, Not Cost per Token
Token prices are easy to compare and mostly misleading. Take one realistic agent job: a long code review with a 200,000-token context (180,000 of it cached), producing 15,000 tokens of output.
- GPT-6 Sol: about $0.23 for that job.
- Claude Opus 5.5: about $0.42 -- roughly 1.8x Sol, with a similar result profile.
- GPT-6 Astra (assuming the same 90% cache discount): about $1.13 -- 5x Sol for the same job.
- GPT-6 Luna on the cheap tier: about $0.01.
Run 100 of those jobs a month and the range is $1 to $113 for comparable, mergeable work. That spread -- not a two-point benchmark gap -- is what decides whether you automate something at all.
What This Changes If You Are a One-Person Business
You do not need a fleet of models. You need three routing rules:
- Cheap tier for volume. Luna-class pricing ($0.10 / $0.50) is for extraction, classification, summarising, tagging, first drafts and anything you will review anyway. At a penny a task it is cheaper than the time you spend deciding whether to run it.
- Mid tier for daily work. Sol, Sonnet 5 and Opus 5.5 are for the work that ships: feature code, code review, document drafting, research synthesis. This is where most of your hours now belong, because the cost per task is cents, not dollars.
- Frontier tier for hard one-shots. Astra and Fable 5.1 still win where a wrong answer is expensive: architecture calls, deep debugging, long migrations, anything you cannot easily verify. Pay 5x on the 5% of jobs that need it.
Two habits make the routing work. Keep prompts cache-friendly -- a stable prefix, or a 90% discount on cached input reads turns into real money on agent loops; recent changes by both vendors let you change reasoning effort or tools without breaking the cache. And measure cost per completed task, not per token: log what each job actually cost, and let the log choose the model for you, because the cheaper model that needs one extra retry is not cheaper.
The strategic read is simple: the frontier is no longer being sold as intelligence. It is being sold as economics. If your competitor has not repriced their workflows this month, you now have a cost structure they do not.
Key Takeaways
- OpenAI's GPT-6 Sol ($2 / $10) and Luna ($0.10 / $0.50) are 50% cheaper than their GPT-5.6 predecessors, and OpenAI says the prices are permanent.
- Anthropic's Claude Opus 5.5 costs $4 / $20 with 60% cheaper cache reads and a 40% lower cost to run than Opus 5 on typical workloads; Sonnet 5.5 and Haiku 5.5 are weeks away.
- Cost per completed task -- price times tokens times steps -- is the metric both vendors now compete on, and it is the one you should track.
- The same week brought Qwen 4 (including a 27B local model), Hy Image 3.5 at about 2 cents an image, and a US government rebrand of AI to "super intelligence".
- For a solo operator, the win is not picking a winner. It is routing three tiers, keeping prompts cacheable, and logging cost per finished job.
If you want the background on how these price tiers were set up in July, read OpenAI's GPT-5.6 launch and what the three-tier pricing meant, and for the provider-by-provider breakdown see the 2026 LLM API pricing deep dive. Both are useful context for the numbers above.
Watch the roundup this piece is based on: the daily AI news video covering GPT-6 Sol and Luna, Opus 5.5, Qwen 4 and Hy Image 3.5. Primary sources: OpenAI: Introducing GPT-6 Sol and Luna and Anthropic: Introducing Claude Opus 5.5.
