Anthropic and OpenAI Launch Cheaper, More Capable Models
Anthropic and OpenAI announced new models on September 22, 2026, each pairing claimed performance gains with lower prices.
Anthropic introduced Opus 5.5. OpenAI introduced GPT-6 Sol and GPT-6 Luna. Engadget OpenAI's news index lists 'Introducing GPT-6 Sol and Luna' as a Product post dated September 22, 2026. Anthropic's news index lists the Claude Opus 5.5 announcement on the same date.
Opus 5.5 is priced at $4 for input tokens and $20 for output tokens, compared with $5 for input and $25 for output for Opus 5. Engadget Input tokens are text fed into the model, output tokens are text it generates. That is a 20% reduction in list API price, the published price developers pay. In separate materials Anthropic says Claude Opus 5.5 costs 40% less to run than Opus 5 and performs at the level of Claude Fable 5.1 on most work. Opus 5.5 is available to developers through Claude, Amazon Web Services, Google Cloud and Microsoft Azure. Engadget
On behavior, Anthropic claims Opus 5.5 attempted to circumvent boundaries around 85 percent less often than past models. Engadget On capability, Anthropic's benchmarks claim Opus 5.5 scored better at agentic coding than GPT-6 Astra on Terminal-Bench 4.0 and FrontierCode v1.1 (Main). Agentic coding means an AI working through multi-step coding jobs with tools, not just answering one prompt. Those are vendor-run numbers. They have not been independently reproduced in the verified facts. Both tests focus on terminal-centric, multi-step coding tasks rather than single-turn completion.
Price, performance and guardrails
OpenAI says GPT-6 Sol and GPT-6 Luna were trained using similar methods to GPT-6 Astra. Engadget Pricing splits the line into a mid-tier workhorse and a low-cost high-volume option. GPT-6 Sol is priced at $2 for input tokens and $10 for output tokens. GPT-6 Luna is priced at $0.10 for input tokens and $0.50 for output tokens.
OpenAI says GPT-6 Sol makes about half as many factual mistakes as its predecessor. The verified facts do not specify the evaluation set behind that claim. For teams running retrieval-augmented generation, where the model pulls in outside documents before answering, tool use, or long-context summarization, factuality rate matters more than raw benchmark peaks. It directly affects verification overhead, retry logic, and human review cost. A halving, if it holds across production workloads, changes inference economics, the cost of running models for users, even before price cuts are factored in.
The pricing structure points to segmentation by workload. Luna pricing sits two orders of magnitude below Opus-class pricing. Common uses at that level include classification, routing, distillation supervision, bulk summarization, and agentic sub-tasks where a larger model acts as planner and smaller models act as executors, much like a senior developer setting the plan while assistants handle routine steps. Sol sits between Luna and Opus 5.5. It is cheap enough for high-throughput coding assistance and enterprise search, while retaining the GPT-6 training lineage.
Where the models ship
Access paths differ. Anthropic is pushing through first-party and hyperscaler channels at once: Claude plus AWS, Google Cloud and Azure. That fits its enterprise pattern of meeting workloads inside existing cloud commits, VPC controls, and procurement vehicles. VPC controls are private network and permission settings inside a company cloud account. It also keeps inference close to customer data gravity, where company data already sits.
OpenAI made GPT-6 Sol and GPT-6 Luna available in ChatGPT Work and Codex for Plus, Pro, Business and Enterprise customers, with Free users and Go subscribers getting only GPT-6 Luna in the desktop app. Engadget Paid tiers get both models for work and coding flows. The free and entry tier funnels to Luna on desktop, which limits expensive inference while seeding distribution.
The launch history behind both includes many steps. Anthropic announced the Claude 3 family on March 4, 2024, Claude Opus 4 on May 22, 2025, Claude Opus 4.5 on November 24, 2025, Claude Opus 4.8 on May 28, 2026, and Claude Opus 5 on July 24, 2026. Anthropic states Claude Opus 4 leads on SWE-bench at 72.5% and Terminal-bench at 43.2%, that Claude Opus 4.8 can work at 2.5x the speed, and that Claude Opus 5 is the strongest Opus model it has tested on its trading benchmark. It announced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. OpenAI's prior price-performance steps include GPT-4o mini on July 18, 2024, described as its most cost-efficient small model and more than 60% cheaper than GPT-3.5 Turbo, and GPT-5.6 Luna on July 30, 2026, described as its fastest and most affordable model at 80% less cost. It also previewed GPT-5.6 Sol as a next-generation model and said GPT-5.6 Terra has competitive performance to GPT-5.5 while being 2x cheaper.
Immediate context helps explain timing. CNBC reported on September 18, 2026 that Anthropic and OpenAI were hunting for smaller AI data center deals in a race to deploy AI capacity. Reuters reported on September 18, 2026 that Anthropic was considering rolling out a new AI model to counter OpenAI's momentum since its launch of GPT-6 Astra. OpenAI launched GPT-6 Astra, according to that same reporting.
The broader context here is a shift from pretraining scale to inference efficiency as the main lever. Both labs now announce price cuts alongside capability claims. That was not always the pattern. Earlier flagship launches emphasized benchmark leadership first, with cost optimization arriving later in mini or distilled variants. Sol, Luna and Opus 5.5 collapse that sequence. Lower cost is part of the launch message.
In my view, tech teams should read the announcements less as a model horse race and more as an operations signal. When vendors cut list prices 20% to 80% while claiming fewer guardrail evasions and fewer factual errors, they are optimizing for agentic deployment at volume. The constraints move to orchestration, eval harnesses, cache hit rates, and incident response when agents act on stale or incorrect outputs.
Worth flagging is that vendor benchmarks on Terminal-Bench 4.0, FrontierCode v1.1, SWE-bench and internal trading tasks are useful for triage, but they do not replace workload-specific evals with your repositories, permissions and data.
What this enables, if the claims transfer, is straightforward. Cheaper, steadier models make it economical to keep agents running longer, to fan out more parallel tool calls, and to add verification passes that were previously too expensive. That favors architectures with a strong planner, cheap executors, and explicit checkers. Over the long arc, that is how general-purpose systems become dependable infrastructure. The risks around misuse persist, as Anthropic's September 2026 threat intelligence report notes in covering seven areas of harm from December 2025 to August 2026, but the direction of travel is toward more capable automation at lower marginal cost.
