GPT-6 Sol and Luna Halve Token Prices: The 2026 Model Price War

Searches for “gpt-6” ran at 2,900 a month in February 2026 and hit 18,100 in August: roughly +86% across the last three months against the prior three. People were not reading about a model. They were waiting for one.

On September 22, 2026, they got two, plus a third from Anthropic, inside about ninety minutes. Anthropic shipped Claude Opus 5.5 at 16:29 UTC. OpenAI followed with GPT-6 Sol and GPT-6 Luna shortly after. The headline is not that frontier models got smarter. The headline is that the price of processing a token at that quality tier fell by half, in an afternoon, and one provider set a date to raise its prices back up.

What actually launched on September 22

Claude Opus 5.5 is Anthropic’s first model in a new 5.5 family. It lists at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. That is a 20% cut on both sides. Anthropic also cut cache reads by 60%, and says Opus 5.5 costs 40% less to serve than Opus 5 on typical workloads. Context window is 1M tokens with a 128K maximum output.

GPT-6 Sol lists at $2 input and $10 output per million. GPT-6 Luna lists at $0.10 input and $0.50 output. Both are exactly half their GPT-5.6 equivalents. Cache reads are $0.20 per million on Sol and $0.01 on Luna, a 90% discount with a 25% premium on cache writes.

Here is the whole mid-tier landscape on the day, in dollars per million tokens.

Model	Input	Cached input	Output
GPT-6 Luna	$0.10	$0.01	$0.50
GPT-5.6 Luna	$0.20	$0.02	$1.20
GPT-6 Sol	$2	$0.20	$10
Grok 4.7	$2	$0.50	$6
Claude Opus 5.5	$4	$0.20	$20
GPT-5.6 Sol	$4	$0.40	$20
Claude Fable 5.1	$10	$0.25	$50
GPT-6 Astra	$10	$1	$50

Read the table twice, because the ordering is not stable. GPT-6 Sol now costs the same as GPT-5.6 Sol did, which is exactly what OpenAI wants you to notice. It also prices below Claude Opus 5.5 on input and output while claiming comparable positioning, which is the pressure Anthropic was responding to before OpenAI even announced.

The part most coverage will skip: GPT-5.6 prices are going up

GPT-5.6 has a scheduled 25% price increase in November 2026. That detail reframes the entire comparison. GPT-6 is half the price of GPT-5.6 today, and it will be roughly 60% of the price of GPT-5.6 after November. Anyone benchmarking “GPT-6 versus GPT-5.6” on current rates is measuring against a promotional floor that is about to be removed.

If you run a workload on GPT-5.6 Terra or Sol today and you have not planned a migration, your November invoice is the one to model. The GPT-6 line appears designed to make that migration the obvious choice rather than a forced one. That is a better way to move users than a deprecation notice, and cheaper than one too.

Cache reads are where a heavy user actually wins

The visible rate cards halved. The less visible change is the cache read price, and for agentic work that is the line that decides your bill.

In a long coding session or a multi-turn agent loop, the overwhelming majority of input tokens you send are not new. They are the same system prompt, the same file context, the same conversation history, re-sent on every single turn. Providers charge those at the cached rate. When 90% or more of your input is cached, a 60% cut to the cache read price is not a rounding error. It is the difference between a $400 month and a $150 month on identical work.

Anthropic cut Opus 5.5 cache reads by 60%. OpenAI prices GPT-6 Sol cached input at $0.20 per million, a 90% discount off the $2 list. If your workflow is one long conversation with a large stable prefix, your effective cost per turn is now governed by the cached number, not the headline number. Fix your prompt structure so the stable prefix sits first and never changes. Move volatile content to the end. That single habit now buys more than any model switch.

Benchmark reality: half the price, mixed intelligence

Artificial Analysis measured the gap between “cheaper” and “better,” and it is worth taking seriously.

Cost per task on their Intelligence Index fell about 50%: GPT-6 Sol (max) runs at $1.06 per task against $1.99 for GPT-5.6 Sol (max), and GPT-6 Luna (max) at $0.07 against $0.18. But the intelligence score itself stayed roughly level. Sol gained 2 points on the Coding Agent Index, with Terminal-Bench 4.0 moving from 37% to 43%. Luna lost 2 points, dropping to 41, with regressions in SWE-Atlas-QnA and DeepSWE.

Hallucination improved substantially. Sol’s hallucination rate fell from 92% to 60% in their knowledge benchmark, and Luna’s from 93% to 77%. The mechanism is worth knowing: Sol answers fewer questions overall, attempting 83% versus 99%, so its accuracy on what it does answer also dropped 5 points. A model that declines more is not the same as a model that knows more.

And the knowledge work results went the other way. In GDPval-AA v2.1, which spans economically valuable tasks across 44 occupations, Sol dropped about 100 Elo points and Luna about 75. Luna also lost roughly 45 points on a multi-week project evaluation. The reported cause is shorter deliverables that omit required elements: cheaper outputs that pass a correctness check while quietly failing a completeness one.

So the honest summary is: same intelligence tier, measurably cheaper, less prone to confident fabrication, and slightly worse at producing complete deliverables. Which of those you weight depends entirely on what you are paying for.

What this means if you are paying $300 or more a month

Four moves, in order of payoff.

Recalculate on cache reads, not list prices. Pull your actual usage report and separate cached from uncached input. If the cached share is high, your real reduction this month is closer to the cache cut than the headline cut. If you are on a subscription plan with no visible cache line item, this is your argument for moving that workload to the API.

Plan your November migration now. Write down every model ID your code and agents reference. Anything pointing at a GPT-5.6 endpoint has a scheduled 25% increase attached. Migrating deliberately in October is cheaper than debugging a bill in December.

Re-run your own evaluation, not the leaderboard. The GDPval regression means general benchmarks will not catch a drop in deliverable completeness. Take ten real tasks you ran last month, run them on GPT-6 Sol and Claude Opus 5.5, and check whether the outputs still contain everything you needed. Cost per correct and complete task is the only number that matters.

Watch the effort setting, especially on Opus 5.5. The “max” thinking level has a 128K output ceiling and it can hit it while still reasoning, returning nothing useful. One documented attempt at a simple SVG generation spent its entire output budget deliberating and failed twice, at a cost of $2.56 per attempt and nearly twenty minutes each. For most work, medium or high effort is the sane default: max effort is a setting that can cost more than it produces.

The cross-provider angle nobody is pricing yet

Look at the bottom of the table. Claude Fable 5.1 and GPT-6 Astra both sit at $10 input and $50 output, ten times the input price of GPT-6 Sol. The price war so far is a mid-tier war. The frontier tier has not moved.

That gap is where routing decisions get interesting. If GPT-6 Sol now does what GPT-5.6 Sol did at half the cost, the correct architecture for a heavy user is a two-tier router: send the routine 80% of traffic to Sol or Luna class models, and reserve the $10/$50 tier for the genuinely hard 20% where the extra intelligence actually changes the outcome. Providers are not going to build that router for you, because every token that stays on the expensive model is revenue.

Anthropic has said Sonnet 5.5 and Haiku 5.5 are coming soon. Current Haiku 4.5 sits at $1 input and $5 output, while GPT-6 Luna is a tenth of that input price. If the small-model tier is where your agent’s inner loop runs, wait for that announcement before you commit, then compare with GPT-6 Luna on your own tasks rather than on the tier label.

The pattern across 2026 is consistent: prices fall in steps, at the middle of the range, announced in competitive bursts, and never announced with a reduction at the top. Your job is not to celebrate the cut. It is to make sure the token mix you actually send lands on the tier that just got cheap, before the next one moves.