OpenAI and Anthropic Both Cut Frontier Prices on the Same Day — Here's What the 50% Headline Actually Means

OpenAI shipped GPT-6 Sol and Luna at 50% cuts over its own promotional pricing, Anthropic shipped Opus 5.5 with a 20% list cut and a 60% cache-read cut, all within hours of each other. The headline 50% is measured against the wrong baseline. The real story is the cache-read cut and the shorter deliverables.

September 23, 2026 – OpenAI shipped GPT-6 Sol and Luna on September 22. Anthropic shipped Claude Opus 5.5 within hours. Both were pitched as frontier price cuts. Both companies confirmed to the press that the cuts are permanent, not promotional. The 50% headline OpenAI used for Sol and Luna, and the 40% headline Anthropic used for Opus 5.5, are both real. Neither one is what an enterprise CFO will see on the next invoice.

Model Input (per 1M tokens) Output (per 1M tokens) Cache reads Cache writes Notes
GPT-6 Sol $2.00 (was $4.00 promo) $10.00 (was $20.00 promo) 90% discount 25% premium Permanent price confirmed by OpenAI to VentureBeat.
GPT-6 Luna $0.10 (was $0.20 promo) $0.50 (was $1.20 promo) 90% discount 25% premium Luna is the high-volume tier; 50% input / 58.3% output cut.
Claude Opus 5.5 $4.00 (was $5.00 list) $20.00 (was $25.00 list) $0.20 (was $0.50) $5.00 (was $6.25) 20% list cut on tokens; 60% cut on cache reads.

The OpenAI “50% off” number is computed against GPT-5.6’s promotional pricing, not the list. The Anthropic “40% off” number is a compound: 20% off list plus “fewer tokens to finish tasks” – Anthropic has not specified which workloads that token reduction applies to. Both numbers are marketing frames built on top of real cuts. The real numbers, in the table above, are the workhorse tier parity: GPT-6 Sol lands at exactly $2/$10, identical to Claude Sonnet 5 (Anthropic made Sonnet 5’s $2/$10 introductory price permanent in August). Closed-weights frontier has converged on a price-per-token schedule: $2/$10 for the workhorse tier, $4/$20 for the premium tier, sub-$1 for the high-volume tier.

The baseline trick

OpenAI’s own launch page describes the cuts as 50% relative to GPT-5.6. The VentureBeat story carries a quoted OpenAI spokesperson confirming the prices are permanent. What the launch page does not say is that GPT-5.6’s list price – the price an enterprise paid without promotional credits – was never $4/$20 for any customer who used the platform at scale. The promotional pricing was the actual pricing for most production workloads. GPT-6 Sol at $2/$10 is a real cut, but it is a cut from promotional pricing, not from list. For workloads that were already on list pricing, the cut is closer to 30%. For workloads on promotional credits, the cut is closer to 50%. The headline number is the bigger of those two, and it is the one OpenAI put in the launch.

Anthropic’s framing is more disciplined. Opus 5.5 is $4/$20 against Opus 5’s $5/$25 list – that is a straight 20% list cut. The “40% less on typical workloads” claim comes from a separate Anthropic materials page and combines the list cut with token-efficiency improvements on common agentic tasks. The compound claim is real, but the token-efficiency half is not a price cut – it is a model-quality claim. If you are budgeting Opus 5.5 against Opus 5 on a fixed workload, your cost will drop somewhere between 20% and 40% depending on how much the model has improved at doing the work. Anthropic has not published a workload distribution, so the right read is “20% guaranteed, up to 40% if the token-efficiency claim holds for your specific tasks.”

The cache-read cut is the real agentic story

Anthropic’s own materials describe cache reads as making up most of the cost of coding agents. A 60% cache-read cut, on top of a 20% list cut, with cache writes down 20% as well, means an Opus 5.5 agent loop is materially cheaper than the headline 20% list cut suggests. For workloads that hit Anthropic’s prompt cache on most turns – which is most production coding-agent workloads – the effective per-task cost drop is in the 35-45% range, not the 20% range. That is the number to budget against.

OpenAI did not move cache pricing on Sol or Luna. GPT-6 Sol and Luna keep the same 90% cache-read discount and 25% cache-write premium that GPT-5.6 had. OpenAI’s pricing model assumes you cache at the prompt layer if you want the discount; Anthropic’s pricing model gives you the discount on the cached portion automatically. For agent loops with high cache hit rates, Anthropic’s effective per-token cost on cached tokens is now lower than OpenAI’s on the same workload shape.

The per-task economics from Artificial Analysis

Artificial Analysis’s Intelligence Index v4.3 – the index that weights coding, reasoning, knowledge work, and agentic benchmarks into a single per-task cost figure – puts GPT-6 Sol at $1.06 per task, down from GPT-5.6 Sol’s $1.99. That is roughly 47% less per task. But the Intelligence Index score for Sol is “level with GPT-5.6” – the cost drop is the price cut alone, not token efficiency. Both models use slightly more output tokens per task than their predecessors: Sol runs at 31k output tokens per task versus GPT-5.6 Sol’s 29k. Luna runs at 51k versus 41k.

The per-task cost story is real. The “level intelligence” half is also real. What is not real is the assumption that you get the cost drop without the token-usage bump. Budget the 47% cost drop, but budget slightly more tokens per task as well.

The regressions nobody is putting in the headline

Coding Agent Index for GPT-6 Sol (max effort) is 57, up 2 from GPT-5.6 Sol. That is an improvement. Coding Agent Index for GPT-6 Luna (max effort) is 41, down 2 from GPT-5.6 Luna. That is a regression – specifically, lower scores on SWE-Atlas-QnA (44% versus 49%) and DeepSWE v1.1 (64% versus 66%). Luna is the cheaper model, and it is the one that got worse on coding-agent benchmarks. The headline 50% applies to both models; the benchmark regression applies to only one.

Knowledge work is where the cheaper-but-shorter trade-off shows up. GPT-6 Sol drops approximately 100 Elo points on GDPval-AA v2.1 – the index that grades model output against rubrics for 44 real-world occupations. GPT-6 Luna drops approximately 75 Elo. Luna drops approximately 45 Elo on AA-Briefcase v1.1, the multi-week knowledge-work index. The Artificial Analysis team’s manual inspection of hundreds of outputs landed on a specific finding: “regressions tend to be driven by reduced presentation quality and deliverables that omit rubric elements.” Cheaper, but deliverables shorter.

The one unambiguous win is hallucination rate. GPT-6 Sol cuts hallucination rate from 92% to 60% on AA-Omniscience. GPT-6 Luna cuts it from 93% to 77%. Sol achieves this by declining to answer more often – 83% of attempts versus 99% for GPT-5.6 Sol – and accuracy drops 5 points in the process (59% to 54%). Luna’s accuracy is broadly unchanged (44% versus 43%). The hallucination rate drop is real. The accuracy trade-off is also real.

Anthropic’s benchmark margins

Anthropic’s Opus 5.5 launch puts the model at 66.4% on Terminal-Bench 4.0 at xhigh effort, with a standard error of plus or minus 2.6 points. That puts Opus 5.5 ahead of OpenAI’s GPT-6 Astra at 57.9% (high effort) and Anthropic’s own Fable 5.1 at 55.8%. The 8.6-point gap over GPT-6 Astra is the headline. The plus-or-minus-2.6 standard error is the caveat – Anthropic itself has cautioned that benchmark margins have become a less reliable guide at this level of capability. The 8.6-point gap is within the margin noise they flag.

The independent framing on OpenAI’s side: DeepSWE v1.1, GPT-6 Sol at max effort scores 68.8%, within 1.1 percentage points of Claude Fable 5’s highest score of 69.9% at xhigh effort. That is OpenAI’s case for “good enough at half the price” – and it is a fair case. The 1.1-point gap is closer than the headline Terminal-Bench 4.0 number suggests, but the price gap is real.

The structural read

When both frontier labs ship workhorse-tier parity pricing on the same day, and one of them (Anthropic) publicly scraps its own 5-hour usage caps on Opus 5.5, that is a competitive signal – not a price story. CNBC framed the dual launch as the first release since the call for a coordinated slowdown. Both companies broke whatever pacing alignment they had during the prior month’s slowdown debate (see TopClanker’s September 16 coverage of the five frontier launches and the news4jax AP-wire piece from September 16 on the slowdown debate).

The structural read is that closed-weights is defending the workhorse tier against open-weight pressure. The prior 60 days produced Xiaomi MiMo-V2.6-Pro at a fraction of GPT-5.6 cost (TopClanker, September 22), Qwen-Image-2.1 at 7B parameters in research-license form (TopClanker, September 21), and Grok 4.7 as the previous closed-weights frontier benchmark. The closed-weights answer is to halve the workhorse price and make the cache cheaper. The routing economy (see TopClanker’s August 18 piece on the Stripe-OpenRouter acquisition) is the surface where this price competition is going to land first – OpenRouter routes by cost-per-task, and the cache-read cut changes the per-task economics for every agent loop on the platform.

The practical takeaway is to budget the price cuts against your actual workload shape. If you are on Anthropic’s prompt cache, the effective drop is in the 35-45% range. If you are on OpenAI’s promotional pricing, the drop is close to 50%. If you are on open-list pricing, the drop is closer to 20-30%. The headline number is the upper bound, not the floor. And budget slightly more tokens per task on Sol and Luna, because both models use a bit more output than their predecessors. The cost drop is real. So is the deliverable-length drop.

Sources