OpenAI and Anthropic Slash Model Costs as AI Price War Moves to Performance per Dollar

The new frontier is not just smarter AI, but cheaper AI

OpenAI and Anthropic have both moved the AI market’s center of gravity from raw model capability toward performance per dollar. OpenAI introduced GPT-6 Sol and GPT-6 Luna as lower-cost members of the GPT-6 family, saying the models bring advances from GPT-6 Astra into faster and more affordable tiers and cut API prices for Sol and Luna by 50% versus GPT-5.6 promotional pricing [1]. Anthropic, meanwhile, launched Claude Opus 5.5 as the first model in its Claude 5.5 family, saying it reaches Claude Fable 5.1-level performance on most work while costing 40% less to run than Opus 5 on typical workloads [2].

The timing matters. Reporting on the launches described Anthropic’s Opus 5.5 release and OpenAI’s Sol-and-Luna response as arriving within roughly an hour of each other, underscoring how quickly frontier labs are now counter-positioning around pricing, coding agents and enterprise deployment [3]. The result is not simply a discount cycle. It is a strategic shift: frontier intelligence is being packaged into more granular tiers, where enterprises can choose whether they need maximum capability, maximum volume, or a balance of both.

OpenAI pushes Astra gains into cheaper Sol and Luna tiers

OpenAI’s pitch is that GPT-6 Astra remains its most capable model, but that many real-world workloads do not need the full flagship every time. GPT-6 Sol is positioned as the more capable everyday workhorse, while GPT-6 Luna is the high-volume, lower-cost option for faster and more routine tasks [1]. In the company’s pricing table, GPT-6 Sol drops from $4 to $2 per million input tokens and from $20 to $10 per million output tokens compared with GPT-5.6 Sol; GPT-6 Luna falls from $0.20 to $0.10 per million input tokens and from $1.20 to $0.50 per million output tokens [1].

That framing is important for developers because token rates are only one side of the bill. Agentic applications often reuse large amounts of context across multiple tool calls, code edits or review loops. OpenAI says improved prompt caching for GPT-6 can help agents reuse more context and obtain 90% discounts on cached input-token reads, which could lower the cost of long conversations and software-agent workflows beyond the headline token-price cut [1].

OpenAI is also tying price cuts to accuracy claims. It says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol on an internal factuality evaluation based on de-identified ChatGPT conversations in which users had flagged prior model errors [1]. MacRumors’ launch coverage similarly summarized Sol as making “about half as many mistakes” as its predecessor, while noting that Sol and Luna remain below Astra as OpenAI’s overall top model [5]. For buyers, that means OpenAI is trying to make the mid-tier feel less like a compromise: lower costs, better reliability, and enough capability for more production workloads.

Anthropic cuts Opus pricing while keeping a premium posture

Anthropic’s Claude Opus 5.5 takes a different route. Instead of introducing a bargain tier, Anthropic is preserving Opus as a premium model while making it cheaper and more efficient. Anthropic lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, with cache reads falling to $0.20 per million tokens [2]. The company says the broader effect is a 40% reduction in typical workload costs because Opus 5.5 uses fewer tokens per task as well as lower per-token prices [2].

This explains why coverage of the price cut may appear to differ. Reuters, via MarketScreener, reported that Opus 5.5 is priced 20% below Opus 5 on input and output token rates, while also noting Anthropic’s claim that the model costs 40% less to run than its predecessor [4]. SiliconANGLE likewise highlighted both comparison bases: a 20% cut in the flagship token price and a 40% lower cost on typical workloads [3]. In other words, the 20% number describes list-token pricing; the 40% number reflects task-level economics, including efficiency.

Anthropic is also emphasizing safety and enterprise readiness. The company says Opus 5.5 was tested before release by outside evaluators including Frontier Design and METR, and it ships with safeguards Anthropic says were developed for its most capable systems [2]. The Reuters report also noted that Anthropic is expanding programs for vetted cybersecurity and life-sciences researchers, reinforcing the company’s strategy of coupling broader model access with controls around sensitive capabilities [4].

Why the “cost per task” argument matters

The clearest lesson from the dual launch is that API price lists are no longer enough to compare frontier models. A model that charges more per token can be cheaper per completed task if it uses fewer tokens, fewer steps or fewer retries. Conversely, a cheaper model can become expensive if it needs repeated attempts or heavy orchestration to reach the same result.

OpenAI makes this argument with Sol and Luna by presenting benchmarks in cost-per-task terms, including claims that GPT-6 Sol can reach or approach more expensive competitors on several coding, computer-use and professional-work evaluations at much lower cost [1]. Anthropic makes the same kind of argument from the opposite direction: Opus 5.5 remains premium-priced versus Sol and especially Luna, but Anthropic says it reduces total workload cost through stronger performance, faster output and lower token use [2].

For enterprise customers, that is the metric that will matter most. Procurement teams care less about which model wins a single benchmark than about how much it costs to resolve a customer ticket, migrate a codebase, audit a contract repository, or run a finance workflow. If a new model reduces both error rates and iteration costs, it can expand the set of tasks where AI deployment is economically rational.

The competitive pressure is now immediate

The launches also show how little room vendors have to let pricing narratives drift. OpenAI compared GPT-6 Sol with earlier Anthropic models in some of its materials, but Anthropic’s same-day Opus 5.5 release complicated that comparison by adding a stronger and cheaper Claude model to the market [5]. MacRumors noted that Anthropic’s new Opus 5.5 outperforms Astra on some coding and knowledge-work benchmarks, while OpenAI’s Sol and Luna undercut Anthropic on token price [5].

That creates a split market. OpenAI can argue that Sol and Luna make high-capability AI more accessible at scale, especially for developers building cost-sensitive applications. Anthropic can argue that Opus 5.5 keeps a premium intelligence tier attractive by lowering the total cost of difficult work. Both arguments can be true, depending on whether a customer’s bottleneck is raw intelligence, latency, reliability, context reuse or budget.

What developers and enterprises should take away

The immediate takeaway is that model selection is becoming more operational. Developers will need to benchmark models against their own workloads rather than rely on generic leaderboards. For routine extraction, summarization, classification and internal support tools, GPT-6 Luna’s very low token rates may be compelling [1]. For coding agents, technical research and longer multi-step tasks, GPT-6 Sol and Claude Opus 5.5 are more direct rivals, with Sol leaning harder into lower list prices and Opus 5.5 leaning into premium performance at lower task cost [1][2].

The deeper takeaway is that the frontier-model business is maturing. Vendors are no longer competing only to announce the most capable model; they are competing to make powerful models usable across more workflows without blowing up inference budgets. That favors customers. If OpenAI and Anthropic continue to compress the cost curve, enterprises can move more AI projects from pilots into production, and developers can afford more experimentation before each product decision becomes a budget debate.

The model war, in short, has entered its efficiency phase. The smartest model still matters. But the model that can be smart often, safely and cheaply may be the one that wins the enterprise deployment cycle.

Sources from the last 72 hours

1. [1]Introducing Claude Opus 5.5 \ AnthropicSep 22, 2026, 4:30 PM UTC
2. [2]Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 modelsSep 22, 2026, 6:45 PM UTC
3. [3]Anthropic unveils Claude Opus 5.5Sep 22, 2026, 4:31 PM UTC
4. [4]OpenAI's New GPT-6 Sol and Luna Models Bring Astra Improvements to Cheaper TiersSep 22, 2026, 8:58 PM UTC
5. [5]Introducing GPT‑6 Sol and LunaSep 22, 2026, 6:16 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.