Competition between OpenAI and Anthropic has entered a new phase. On September 22, 2026, Anthropic introduced Claude Opus 5.5, and OpenAI followed the same day with GPT-6 Sol and GPT-6 Luna. All three models are built around the same commercial proposition: advanced performance at a substantially lower operating cost.
The near-simultaneous launches represent more than another round in the race to build the smartest model. They show that inference prices, token efficiency and the cost of completing a task have become central considerations for businesses deploying artificial intelligence in production. The difference may look small for a low-traffic chatbot. For an AI agent that reads documents, calls tools, writes code, checks its work and repeats the process thousands of times, it can transform the economics of the entire product.
Two launches, one objective
Anthropic described Claude Opus 5.5 as the first model in its new Claude 5.5 family. The company says it performs at roughly the level of Claude Fable 5.1 on most work while costing about 40% less to run than Opus 5 on typical workloads. Standard API pricing has fallen to $4 per million input tokens and $20 per million output tokens, compared with $5 and $25 for Opus 5.
Anthropic’s estimated saving is larger than the 20% reduction in standard token rates because Opus 5.5 is also designed to consume fewer tokens when completing a task. The company says the model generates output more than 30% faster than Opus 5 and required fewer steps, tool calls and output tokens in its own evaluations.
Roughly 90 minutes after Anthropic’s announcement, OpenAI expanded the GPT-6 family with GPT-6 Sol and GPT-6 Luna. Sol is positioned for coding, agents and complex workflows, while Luna targets fast, high-volume tasks. OpenAI says both were trained with methods similar to those used for GPT-6 Astra but optimized to provide greater speed and cost efficiency.
OpenAI priced GPT-6 Sol at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs only $0.10 for input and $0.50 for output. OpenAI describes the new rates as 50% below the promotional prices of the corresponding GPT-5.6 models, although Luna’s output reduction from $1.20 to $0.50 is mathematically slightly greater than 50%.
API prices compared
All prices below are in US dollars per million tokens. The figures cover standard processing; fast modes, batch processing, very long contexts and third-party cloud platforms may have different rates.
| Model | Positioning | Input | Cached input | Output |
|---|---|---|---|---|
| GPT-6 Luna | Fast, high-volume tasks | $0.10 | $0.01 | $0.50 |
| GPT-6 Sol | Coding and agentic workflows | $2 | $0.20 | $10 |
| Claude Opus 5.5 | Agentic coding and knowledge work | $4 | $0.20 | $20 |
| GPT-6 Astra | OpenAI premium model | $10 | — | $50 |
| Claude Fable 5.1 | Anthropic premium model | $10 | $0.25 | $50 |
The table gives OpenAI a clear advantage on nominal pricing. GPT-6 Sol costs half as much as Claude Opus 5.5 for both standard input and output, while GPT-6 Luna operates in a much lower price category. This is not yet a complete cost comparison, however. A model with cheaper tokens may still cost more per successful result if it writes longer responses, makes more tool calls or needs more attempts.
Token price is not enough
Tokens remain the billing unit, but they are not the economic unit that matters to most organizations. A business does not buy tokens simply to process them. It pays to resolve a support ticket, validate an invoice, prepare a report, modify an application or complete a sales workflow.
The real cost of a task can be expressed as:
Cost per task = input + output + cache reads and writes + tool calls + retries + human or automated validation.
This explains why a low input price does not guarantee the lowest total cost. A coding agent may load hundreds of thousands of tokens from a repository, search several files, run tests and revisit its proposed change. A research agent may perform dozens of searches and analyse the same documents repeatedly. A sales agent may cross several applications before it updates a CRM correctly.
Efficiency matters as much as the price list in these scenarios. If one model completes the task in five steps while another needs ten, the first may be cheaper even when each of its tokens costs more. A model that fails less often also reduces hidden expenditure on reruns, output verification and human intervention.
What cost per task reveals
OpenAI used AutomationBench to illustrate this distinction. The benchmark tests AI agents on end-to-end workflows using 47 tools across sales, marketing, operations, support, finance and human resources. In the published results, GPT-6 Sol at xhigh scored 33.2% at $0.27 per task, while Claude Opus 5 at maximum effort scored 26.9% at a cost about 11.1 times higher.
This comparison supports OpenAI’s efficiency argument, but it requires context. The Anthropic model in the comparison is Opus 5, not the new Opus 5.5 released that day. The results also use particular configurations and effort levels, and no benchmark can reproduce every condition found in a production deployment.
Anthropic makes its economic case differently. It says Opus 5.5 costs 40% less per typical task than Opus 5 because it combines lower rates with reduced token consumption. In Anthropic’s agentic coding evaluations, the model reportedly matched its predecessor’s quality using roughly half the turns, time and output tokens in some workflows.
Neither claim should be treated as a universal verdict. The figures are published by the model providers, and every benchmark emphasizes particular task types. The decisive evaluation for a business remains a test using its own data, tools and acceptance criteria.
Caching becomes a pricing weapon
One of the most important changes is the lower price of prompt caching. This technique allows applications to reuse a stable portion of a prompt — such as system instructions, documentation, internal policies or the context of a software project — without paying the full input rate every time.
A Claude Opus 5.5 cache read costs $0.20 per million tokens, 60% below the $0.50 charged for Opus 5. Anthropic prices a five-minute cache write at $5 per million tokens and a one-hour write at $8, after which qualifying cache reads are much cheaper.
GPT-6 Sol provides cached input at $0.20 per million tokens, or 10% of its standard input price. Luna’s rate is $0.01. OpenAI attributes part of the broader price reduction to improvements in caching and inference efficiency.
The impact can be substantial for agents that repeatedly reuse the same context. Caching does not reduce a bill automatically, however. Applications must preserve identical prompt prefixes, monitor their cache-hit rate and account for the price of writing the cache. A cache created and read only once may provide little or no saving.
Pressure from Chinese and open models
OpenAI and Anthropic are not competing only with each other. Open-weight models and Chinese providers have sharply reduced the market price for capabilities that were considered premium only a short time ago. Moonshot, DeepSeek and Alibaba have introduced models designed to compete on performance, open distribution and cost simultaneously.
OpenAI had already cut GPT-5.6 Luna prices by 80% and GPT-5.6 Terra prices by 20% during the summer of 2026. Anthropic had introduced Opus 5 at half the price of Fable 5 while retaining much of its capability. Those decisions preceded the latest launches and indicate that price pressure is structural rather than a one-off promotion.
Lower-priced alternatives change the buyer’s negotiating position. If a task does not need frontier performance, a business can use a low-cost or open-weight model and reserve premium systems for exceptions. This multi-model architecture makes it harder for any provider to preserve a large premium on capabilities that have become standardized.
Prices cannot fall indefinitely. GPUs, energy, memory, networking and serving capacity remain expensive. State Street’s analysis argues that open models create a ceiling on intelligence prices while scarce compute creates a floor. Competition can compress margins, but inference does not become free.
Implications for developers
For AI application developers, lower prices create room for more testing and iteration. Features that were previously too expensive at scale — analysing every document, classifying every message, automatically checking results or running multiple agents — may become commercially viable.
The cheapest model should not be selected directly from a pricing table. A practical evaluation should measure:
-
Successful task-completion rate, not only the quality of one response.
-
Average cost and 95th-percentile cost for each task.
-
Tokens, tool calls and attempts required.
-
Time to an accepted result, rather than generation speed alone.
-
Cache-hit rate and cache-write expenditure.
-
Frequency of escalation to a stronger model or human operator.
-
Stability after model-version updates.
A mature application can use dynamic routing. Luna or an open-weight model can process classification, extraction and basic summarization. Sol or a model in the Sonnet tier can handle intermediate workflows. Opus 5.5, Fable 5.1 or Astra can be called only when the difficulty, risk or value of the task justifies their price.
This architecture reduces expenditure without sacrificing performance on important cases. It also protects the application against excessive dependence on one provider. To work reliably, it needs a shared layer for prompts, tools, logging, evaluations and model-version control.
Implications for businesses
For an organization buying AI services, the new prices move the conversation from “which model has the highest benchmark?” to “what does an accepted result cost?”. This is important for budgeting because the API bill represents only part of the total expenditure.
Total cost of ownership includes integration, observability, security, data governance, evaluations, support and employee time. A cheaper model may increase operating expenditure if it produces more errors. Conversely, a premium model may be justified when it reduces human intervention in a high-value process.
Businesses can benefit from the price war through three immediate actions:
-
Re-evaluate existing workloads. A recently released model may deliver comparable quality at a substantially lower cost than one selected only a few months earlier.
-
Separate tasks by difficulty and risk. Not every request should be sent to the most capable model.
-
Negotiate around actual usage. Volume, caching, batch processing, rate limits and regional availability may matter more than the headline public rate.
How to access the models
GPT-6 Sol and GPT-6 Luna are available through the OpenAI API under the identifiers gpt-6-sol and gpt-6-luna. At launch, OpenAI also began rolling them out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, while Luna became available in the desktop app for Free and Go users.
Claude Opus 5.5 is available through the Claude Platform under claude-opus-5-5. Anthropic also announced availability through Amazon Web Services, Google Cloud and Microsoft Azure, allowing organizations to integrate it through an existing cloud environment.
For current figures, consult the official OpenAI pricing page and Anthropic pricing documentation, because promotional rates, long-context multipliers and processing modes can change.
Frequently asked questions
Which is cheaper: GPT-6 Sol or Claude Opus 5.5?
At standard API rates, GPT-6 Sol is cheaper: $2 per million input tokens and $10 for output, compared with $4 and $20 for Claude Opus 5.5. Both charge $0.20 per million tokens for cached input reads.
Is GPT-6 Luna the best option for every high-volume application?
Not necessarily. Its very low rate makes it attractive for classification, extraction, moderation, text processing and other repetitive operations, but the decision must be validated against success rate, latency and retries. A model that costs less per token can still cost more per correct outcome.
Why does Anthropic say Opus 5.5 is 40% cheaper when token prices fell by 20%?
The estimate combines two effects: a lower token rate and a claim that the model uses fewer tokens and steps on typical work. The 40% reduction is based on Anthropic’s evaluations and is not a guarantee for every workload.
Are benchmarks enough to choose a model?
No. Benchmarks provide a useful starting point, but tool configuration, prompts, data, reasoning effort and success criteria can all change the outcome. An internal test built from real tasks and recorded costs is more relevant than a single public score.
The market’s next phase
The OpenAI–Anthropic price war shows that advanced intelligence is becoming an increasingly competitive commodity. Differentiation will not disappear, but it is moving toward efficiency, latency, reliability, tool integration, caching, safety and the ability to complete real work.
Lower prices are good news for users. They place pressure on provider margins and force AI labs to convert infrastructure and model-efficiency gains into visible customer savings. For developers and businesses, the practical rule is straightforward: the right model is not necessarily the smartest one or the model with the cheapest tokens. It is the one that delivers an accepted result at the best total cost.
