OpenAI and Anthropic both cut coding-model prices the same week, and agent loops are where it lands

On September 22 the two labs that set the price of frontier inference both moved it down. OpenAI released GPT-6 Sol and GPT-6 Luna and cut their API rates by about half. Anthropic shipped Claude Opus 5.5 and priced it below the model it replaces. The version numbers and benchmark tables will get the attention, but the durable change for anyone building on these models is quieter: the cost of the token-heavy agent loops both companies are now selling to just dropped, and it dropped most where those loops spend the most.

OpenAI pushed capability down a tier

GPT-6 Sol lists at $2 per million input tokens and $10 per million output, half of what GPT-5.6 Sol charged. GPT-6 Luna drops to $0.10 and $0.50, roughly a 50 to 58 percent cut. OpenAI is explicit that these are permanent rates, not promotional pricing that expires. Sol is aimed at the repeatable work developers already hand to a model: building features, reviewing code, debugging, and working through data. Luna is the high-volume option for narrower jobs like extraction, summarization, and answering straightforward questions.

The positioning is the interesting part. OpenAI is not pricing a new flagship, it is pushing near-flagship capability into a mid tier and undercutting the field on cost per token. Its own comparison names the target directly: on AutomationBench 1.0.6, OpenAI reports GPT-6 Sol at high effort scoring 33.2 percent at $0.27 per task against Claude Opus 5 at 26.9 percent while costing about eleven times as much per task. That is a vendor-run benchmark against Anthropic’s prior flagship, not the one Anthropic shipped the same day, so read it as a statement of intent rather than a settled result.

Anthropic made its flagship cheaper and faster

Claude Opus 5.5 lists at $4 per million input tokens and $20 per million output, a 20 percent cut from Opus 5, with cache reads down to $0.20 per million, a 60 percent reduction. Anthropic puts the practical savings at about 40 percent on typical workloads and says output generation is more than 30 percent faster. Unlike OpenAI’s move, this is a top-tier model getting cheaper rather than a cheaper tier getting more capable.

Anthropic’s benchmark table leans on agentic coding: it reports Terminal-Bench 4.0 at 66.4 percent against Opus 5’s 52.3 percent, and CursorBench 4.0 at 57.8 percent. The claims worth holding onto are the concrete ones, because they map to time and money at once. Anthropic says one early tester ran a 200,000-line code audit in under three hours where Opus 5 took more than twenty. A task that finishes in a seventh of the wall-clock time on a model that also costs less per token is the combination that actually moves a budget, since the hours and the tokens fall together.

The savings compound inside the loop

Both cuts land hardest on the same workload, and it is not the one-shot prompt. An agent loop re-reads a large, mostly static context on every turn: the system prompt, tool definitions, prior steps, and the files in scope. That re-read is exactly what a prompt cache serves, so a cut to cache-read pricing compounds across a long-running session in a way it never does for a single call. Anthropic’s 60 percent cache-read reduction is aimed straight at that pattern, and OpenAI’s flat 50 percent cut lowers the floor under every turn regardless of caching.

For a team running coding agents at any real volume, the metric that moved is cost per completed task, not headline token price. Agents burn tokens on retries, tool calls, and re-read context, so the per-task bill is what a lower rate and a higher first-pass success rate quietly reshape. Both labs are betting that once the cost of a full agent run drops far enough, workloads a team currently rations become ones it can leave running.

What to watch

The one comparison that reads cleanly is the price line. GPT-6 Sol at $2 and $10 now sits at exactly half of Claude Opus 5.5’s $4 and $20, which is to say Opus 5.5 lands where GPT-6 Sol’s old price was. That looks like a direct undercut, but the two are not the same product: Sol is a mid tier below GPT-6 Astra, and Opus 5.5 is Anthropic’s flagship, so the gap partly reflects where each model sits in its own lineup rather than a like-for-like discount. The real test is production cost per task on a team’s own workloads, not either lab’s table. Over the next quarter the signals worth watching are whether independent evaluations reproduce the agentic-coding gains outside vendor benchmarks, whether the cache-read savings hold up in messy real loops rather than clean ones, and whether OpenAI’s mid-tier pricing pressures Anthropic to move its flagship rate again. Stackmaven’s follow-up will revisit those around December 15.

Sources cited

1. Introducing GPT-6 Sol and Luna (OpenAI) openai.com
2. Claude Opus 5.5 (Anthropic) www.anthropic.com
3. OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more (VentureBeat) venturebeat.com
4. Anthropic releases Claude Opus 5.5 with lower pricing and new safeguards (Unite.AI) www.unite.ai
5. OpenAI's GPT-6 Sol and GPT-6 Luna now available (GitHub Changelog) github.blog