The price war: GPT-6 Sol and Luna at half price, dropped 90 minutes after Claude Opus 5.5

The Price War: GPT-6 Sol and Luna at Half Price, Dropped 90 Minutes After Claude Opus 5.5

🔎 90 minutes. That's all Anthropic got.

September 22, 2026 will go down in AI's short history as the day the price war became total. Anthropic releases Claude Opus 5.5 at $4/$20 per million tokens — already 40% less than Opus 5. The blog post is barely online when OpenAI strikes: GPT-6 Sol and GPT-6 Luna land in the API, Codex, and ChatGPT, at half price.

The numbers are brutal. Sol goes from $4/$20 (the promo rate of GPT-5.6) to $2/$10. Luna drops from $0.20/$1.20 to $0.10/$0.50, a −58% cut on output. And unlike the company's usual habits, an OpenAI spokesperson assures that this pricing is permanent, not promotional.

Why now? Because the frontier race has changed in nature. It's no longer just about who is the smartest, but about who costs the least per completed task. That morning, OpenAI decided to write the new rule of the game: the same level of performance, twice as cheap, launched an hour and a half after the competitor.

The essentials

• GPT-6 Sol: $2/$10 per million tokens (input/output), half the promo price of GPT-5.6 Sol ($4/$20). Pricing presented as permanent (September 2026).
• GPT-6 Luna: $0.10/$0.50, i.e., −58% on output compared to GPT-5.6 Luna ($0.20/$1.20).
• Launched 90 minutes after Claude Opus 5.5 ($4/$20), on the same day. The timing is a message, not a coincidence.
• Half as many factual errors: 4.6% for GPT-6 Sol vs. 8.5% for GPT-5.6 Sol, confirmed by independent measurements from MetricNexus (2026).
• But Sol regresses on two agentic benchmarks: DeepSWE v1.1 (68.8% vs. 72.7%) and OSWorld 2.0 offline (64.4% vs. 66.2%). Migration warrants testing.
• Cost per task in free fall: from −45% to −94% depending on the benchmark (Artificial Analysis, MetricNexus, 2026).
• No GPT-6 Terra. No fine-tuning. 1.05M token context, 128k max output.

Recommended Tools

To decide based on your specific use case, our monthly comparison of the best LLMs is updated with every major release.

Pricing, line by line

Direct answer: GPT-6 Sol costs $2 per input and $10 per output per million tokens. GPT-6 Luna, $0.10 and $0.50. And OpenAI presents these prices as permanent, not as a limited-time promo.

Here's the full breakdown, cross-referenced between MetricNexus, NotesByLex and Codersera (September 2026):

On non-cached input and output, Sol is therefore half the price of Opus 5.5. Luna is literally 40 times cheaper than Opus 5.5 on output. That's the new market floor, and it was set in a single morning.

The terms that change the bill

Three subtleties to know before migrating:

• Prompts over 272K tokens: input and cache billed at double, output at 1.5×. The 1.05M token context comes at a price.
• Batch and Flex: −50%. Fast Mode: ×2. It's up to you to choose between speed and budget.
• No fine-tuning on the GPT-6 family, and Tier 1 rate limits of 500 RPM / 500K TPM (Codersera, 2026).

The IDs to remember: gpt-6-sol and gpt-6-luna. Context of 1.05M tokens, max output 128k. Knowledge cutoff: April 2026 for Sol, May 2026 for Luna.

90 minutes after Opus 5.5: the coincidence that isn't one

Direct answer: no, it's not a coincidence, and OpenAI isn't doing anything to hide it.

Let's reconstruct the day's timeline, as reported by The New Stack and AI Weekly. Anthropic releases Claude Opus 5.5: $4/$20, i.e. 40% less than Opus 5, with 66.4% on Terminal-Bench 4.0 and 81.8% on OSWorld 2.0 to boot, according to Anthropic's official spec sheet.

90 minutes later, OpenAI announces Sol and Luna. The message is readable without glasses: whatever you release, I release something cheaper the same day. This isn't a product launch, it's a calculated counterpunch.

This sequence fits into an escalation that began earlier in the month. GPT-5.6 Sol had already opened the front with an aggressive promotional price — we analyzed it in our article on the GPT-5.6 Sol preview, right at the start of the price war. Then GPT-6 Astra arrived on September 3 at $10/$50, positioned at the very top of the ladder. Sol and Luna round it out at the bottom.

Anthropic, for that matter, is no stranger to this kind of duel: the brand had already come out on top on benchmarks with Claude Opus 4.8 vs GPT-5.5, and the rivalry now shapes the entire release calendar. Our Claude vs ChatGPT comparison details what separates the two ecosystems beyond price.

My take: the spokesperson's word "permanent" should be taken seriously in its letter, not in its spirit. Permanent doesn't mean eternal; it means "not a promo that expires in 60 days." In a market where Anthropic has to respond, expect further price cuts before the end of the year.

Benchmarks: cheaper and more reliable, but not better everywhere

Direct answer: GPT-6 Sol doesn't crush GPT-5.6 Sol. It is significantly cheaper and more factually reliable, but it regresses on some agentic benchmarks. It's a deliberate trade-off, not a free upgrade.

The point everyone remembers: factual errors

OpenAI claims that Sol makes half as many factual errors as its predecessor. MetricNexus (2026) confirms: 4.6% errors for GPT-6 Sol versus 8.5% for GPT-5.6 Sol. This isn't a slogan, it's an independently measured figure.

For any use case where truthfulness matters — writing, customer support, document analysis, monitoring — this is the #1 argument of this launch. An error rate cut in half changes the real cost of human review, not just the API bill.

The point everyone will forget: agentic regressions

Sol regresses on DeepSWE v1.1, an autonomous software engineering benchmark: 68.8% versus 72.7% for GPT-5.6 Sol, with the gap widening to −8.2 points in xhigh mode. Same story on OSWorld 2.0 offline, which measures agents on system tasks: 64.4% versus 66.2%.

In other words, on the hardest agentic tasks, the old model at $4/$20 can still beat the new one at $2/$10. This is exactly the kind of detail that press releases don't put in bold, and that independent benchmarks reveal.

Luna, meanwhile, wins or ties everywhere at maximum effort, with, for example, 66.6% versus 62.2% on DeepSWE (MetricNexus, 2026). A remarkable fact: the little sister at $0.10/$0.50 is more consistent than the big brother on this specific front.

The real tiebreaker: cost per task

Artificial Analysis (2026) puts Sol (max) at $1.06 per task on the Intelligence Index, versus $1.99 for GPT-5.6 Sol max — roughly 50% less. Luna (max) drops to $0.07 per task, roughly 60% less. And on the Agents' Last Exam, Luna goes from $2.57 to $0.15 per task: −94% (MetricNexus, 2026).

Sol shows −58 to −59% cost per task on 4 of the 5 benchmarks measured. The Intelligence Index and Coding Agent Index scores remain stable according to Artificial Analysis, with gains and regressions depending on the evaluation.

My take: OpenAI traded a bit of agentic margin for an explosion in cost efficiency. For complex production code, this means re-testing your workflows before switching over. For everything else, the trade-off is already won.

Availability: ChatGPT Work, Codex, and the gradual rollout

Direct answer: Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu subscribers, with a gradual rollout spread throughout the day.

This gradual rollout is not a technical detail. OpenAI is explicit about it: it serves service stability, to prevent millions of users from switching to the new models all at once (OpenAI Developer Community, 2026). Both models build on the advances of GPT-6 Astra, with optimized caching and inference whose savings are passed back into the price.

For individual users without a subscription, there's a genuine novelty: Free and Go tiers get access to Luna in the desktop app. The best value for money in the proprietary market thus also becomes a free entry point — enough to reshape our selection of the best free LLMs.

In the API, nothing complicated: the IDs gpt-6-sol and gpt-6-luna are all you need, with no fine-tuning available and no special configuration. The simplicity of migration is deliberate — the faster you switch, the more the price war benefits them as much as you.

The cache trap: "half price" doesn't mean your bill is cut in half

Direct answer: the half price only applies to non-cached tokens. When it comes to caching, the price war hasn't spread yet.

The unwelcome detail comes from NotesByLex (2026): GPT-6 Sol's cache reads cost $0.20, exactly like those of Claude Opus 5.5. If your workload relies on long, reused system prefixes — the typical case for agents — the real price gap shrinks to virtually nothing on the cached portion of your requests.

Second subtlety: tokenizers and reasoning lengths differ from one model to another. The token ratio is not the cost-per-task ratio, warn NotesByLex and Artificial Analysis. Comparing $2 and $4 raw is a marketing exercise, not accounting.

GPT-6 cache, enabled by default

Good news nonetheless: caching is enabled by default on the GPT-6 family, with a minimum prefix of 1,024 tokens, a TTL of at least 30 minutes, writes at 1.25× the input price and reads at 10% of that price (NotesByLex, 2026). Artificial Analysis confirms the same 90% reduction on cache reads and a 25% premium on writes.

In practice, this gives a simple rule. If your workflows are agentic and repetitive — the same system context replayed hundreds of times — calculate your cost factoring in the cache, not the list price: that's where the real savings lie. If your requests are short and unique, the raw price applies and Sol is unbeatable. The two scenarios don't tell the same story.

Sol for code, Luna for volume: who should take what

Direct answer: Sol for complex code and demanding agents, Luna for bulk work, Astra for research. The split is clear, own it.

GPT-6 Sol targets complex code. It's the model for refactors, multi-file debugging, and long development sessions. It's available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu subscribers (The New Stack, 2026). Our guide to the best LLMs for coding will be updated with the consolidated benchmarks — keep the DeepSWE regression in mind before migrating a production pipeline.

GPT-6 Luna targets bulk work. Classification, extraction, summarization, moderation, RAG pipelines: anything that runs millions of times a day. At $0.10/$0.50, it literally changes the economies of scale for these use cases — the −94% cost per task on Agents' Last Exam is the most spectacular demonstration of this.

GPT-6 Astra remains the top of the lineup. At $10/$50, released on September 3, it targets research and the problems Sol can't solve. We covered its 64.6% on Terminal-Bench and its science performance in our article on GPT-6 Astra.

If you're building agents, the Sol vs Opus 5.5 question deserves a proper test bench rather than a snap judgment: our feature on the best LLMs for AI agents lays out the testing methodology. And for hallucination-sensitive use cases, the 4.6% vs 8.5% factual error rate weighs more heavily than any price.

A price war that's only just beginning

Direct answer: the frontier input price has been divided by five in three weeks, and nothing suggests the trend is stopping.

Let's do the math. On September 3, GPT-6 Astra is listed at $10/$50. On September 22, Sol arrives at $2/$10. In 19 days, frontier input went from $10 to $2. No technological breakthrough happened in between — it's a business decision, and OpenAI owns it as such: Sol and Luna are presented as faster, cheaper models for high-volume work, with caching and inference savings passed on to customers (OpenAI Developer Community, 2026).

Two pressures explain this aggressiveness. The first comes from Anthropic, which just slashed its own pricing with Opus 5.5. The second comes from open models: MiniMax M3 challenges GPT-5.5 with 1M context and a novel architecture, as we described in our article on MiniMax M3, the Chinese open-weights model that challenges GPT-5.5. When a downloadable model gets close to the frontier, proprietary providers have no weapon left but price.

Note also the transition buffer: GPT-5.6 Sol keeps its promo pricing "at least until November 21, 2026" (The New Stack, 2026). OpenAI is leaving an exit door for those who don't want to migrate immediately — and making sure nobody pays more by accident during the switchover.

My prediction: Anthropic will respond before the end of the quarter, probably on caching or via an intermediate model. In this context, the only rational strategy for a developer is to stay agnostic: architect things so you can swap models by changing a single line, and re-evaluate every month. That's exactly what our monthly comparison of the best LLMs does.

❌ Common Mistakes

Mistake 1: comparing prices token to token

The price per million tokens is not the real cost. Tokenizers and reasoning lengths differ from one model to another, so the price ratio does not reflect the cost-per-task ratio (NotesByLex, 2026). The solution: measure the cost per task on your own workloads, as Artificial Analysis does on its benchmarks.

Mistake 2: migrating agents without re-testing

GPT-6 Sol regresses on DeepSWE v1.1 (−3.9 points, up to −8.2 in xhigh) and on OSWorld 2.0 offline. If your agents write code or operate a system autonomously, run your test suite before changing the model ID. Luna may even prove more stable than Sol on some agent workflows — test both.

Mistake 3: forgetting surcharges and multipliers

Beyond 272K prompt tokens, input is billed at double and output at 1.5×. Fast mode doubles the bill, Batch halves it. A misconfigured pipeline can cost twice the advertised price without anyone noticing until the end of the month.

Mistake 4: taking "permanent" for a contractual commitment

Permanent means "not a limited-time promo", not "guaranteed forever". In a market where prices move twice a month, protect yourself with a provider abstraction in your code, not with a spokesperson's word.

❓ Frequently Asked Questions

Does GPT-6 Sol replace GPT-5.6 Sol?

No, both coexist in the API. GPT-5.6 Sol keeps its promotional pricing of $4/$20 at least until November 21, 2026, which leaves a comfortable transition period. OpenAI has not announced a deprecation date for the 5.6 series at this stage.

Does GPT-6 Terra exist?

No. Unlike the GPT-5.6 family, which included Sol, Terra and Luna, no GPT-6 Terra has been announced with this launch. Only Astra, Sol and Luna make up the GPT-6 family as of today (September 2026), according to The New Stack and Codersera.

Can you use GPT-6 Luna for free?

Yes, partially. Free and Go users have access to Luna in the ChatGPT desktop app. Sol and both models in ChatGPT Work and Codex require a Plus, Pro, Business, Enterprise or Edu subscription. In the API, Luna is billed at $0.10/$0.50 per million tokens.

Is fine-tuning needed on these models?

That's not possible: OpenAI does not support any fine-tuning on the GPT-6 family (Codersera, 2026). Customization goes through prompting, prefix caching and tools, not through training. Tier 1 rate limits are 500 RPM and 500K TPM.

Should you switch from Claude Opus 5.5 to GPT-6 Sol?

Not automatically. Sol is half as expensive on non-cached tokens, but Opus 5.5 claims 66.4% on Terminal-Bench 4.0 and 81.8% on OSWorld 2.0, and Sol regresses on some agentic benchmarks. Test your workloads, compare the real cost per task, then decide.

What is the maximum context of GPT-6 Sol and Luna?

1.05 million tokens in input, with a maximum output of 128k. Note, however: beyond 272K prompt tokens, input and cache pricing is doubled and output is billed 1.5 times higher. Knowledge cutoffs are April 2026 for Sol and May 2026 for Luna.

✅ Conclusion

In 90 minutes, on September 22, 2026, frontier pricing was cut in half — with doubled factual reliability thrown in as a bonus, though a few agent regressions that independent benchmarks don't hide. The price war has become AI's main battlefield, and its big winner, for once, is you. To choose your model this month, head over to our comparison of the best LLMs.