GPT-6 Sol vs Claude Opus 5.5: The Cost of AI Intelligence
OpenAI and Anthropic have released new models almost simultaneously, putting a growing emphasis on a metric that is becoming increasingly important for developers and enterprise AI deployments: the cost required to complete a task.
OpenAI introduced GPT-6 Sol and GPT-6 Luna as lower-cost models derived from the GPT-6 Astra generation, while Anthropic launched Claude Opus 5.5 as a more efficient successor to Opus 5.
Both releases focus heavily on inference efficiency rather than simply maximizing benchmark scores. OpenAI reduced the API prices of GPT-6 Sol and Luna by 50% compared with the promotional pricing of GPT-5.6, while Anthropic says Opus 5.5 reduces typical task costs by approximately 40% and increases output speed by more than 30% compared with Opus 5.
The result is a notable shift in how frontier AI models are being positioned. Instead of evaluating models solely by intelligence or benchmark performance, developers increasingly need to consider the total cost of completing real-world work.
⚡ OpenAI Releases GPT-6 Sol and Luna at Lower Cost #
OpenAI previously introduced GPT-6 Astra as its highest-capability model in the GPT-6 generation. GPT-6 Sol and GPT-6 Luna extend that technology into lower-cost, higher-throughput configurations designed for broader deployment.
The basic strategy is straightforward: Astra serves as the capability frontier, while Sol and Luna bring much of that technology into workloads where latency, throughput, and operating cost matter more.
OpenAI has emphasized that model capability is only one component of enterprise AI economics. For production systems, the more important question can be how much it costs to successfully complete a task.
OpenAI’s announcement:
https://openai.com/index/introducing-gpt-6-sol-and-luna/
GPT-6 Sol and Luna Pricing #
Compared with GPT-5.6 promotional pricing, OpenAI says GPT-6 Sol and GPT-6 Luna reduce API prices by approximately 50%.
- GPT-6 Sol: Input pricing falls from $4 to $2 per million tokens, while output pricing drops from $20 to $10 per million tokens.
- GPT-6 Luna: Input pricing falls from $0.20 to $0.10 per million tokens, while output pricing drops from $1.20 to $0.50 per million tokens.
The lower prices are paired with optimization of caching and inference efficiency, allowing developers to run more workloads within the same compute budget.
Enterprise Workflow Performance #
On AutomationBench, which evaluates enterprise workflows across applications, OpenAI reports that GPT-6 Sol at maximum reasoning effort outperformed Claude Opus 5 at its maximum effort while costing approximately 9% as much per task.
At high reasoning effort, GPT-6 Luna improved by 5.4 percentage points over the previous-generation model while reducing per-task cost by 58%.
OpenAI also reports that GPT-6 Sol surpassed Claude Fable 5.1 on selected evaluations at lower cost and exceeded GPT-6 Astra when Astra was operated at low reasoning effort.
In internal factual-accuracy evaluations based on anonymized real-world conversations where users had previously identified model errors, OpenAI says GPT-6 Sol produced roughly half as many errors as its predecessor, approaching GPT-6 Astra’s reliability while operating at lower cost.
GPT-6 Luna also shows a substantial efficiency advantage. OpenAI reports that at higher reasoning effort, Luna can approach the performance of GPT-5.6 Sol at approximately one-hundredth of the cost.
Coding and Software Engineering #
Software development is another major target for these efficiency improvements.
As coding agents take on longer and more complicated tasks, inference expenditure becomes an important factor for development teams. Lower model prices allow developers to run more experiments, delegate more work to coding agents, and use systems such as Codex across a broader range of workflows.
On FrontierCode, OpenAI reports a significant improvement for GPT-6 Sol compared with GPT-5.6 Sol, with performance reaching the level of Claude Fable 5.1 at xhigh reasoning effort at substantially lower cost.
On DeepSWE v1.1, GPT-6 Sol scored 68.8% at maximum reasoning effort, compared with 69.9% for Claude Fable 5. OpenAI reports that GPT-6 Sol’s per-task cost was approximately 80% lower.
GPT-6 Luna reached 66.6% at maximum reasoning effort, approaching the medium-effort performance of Claude Opus 5 and Fable 5.
In the same comparisons, OpenAI reports that Luna reduced per-task cost by 93% compared with Opus 5 and by 96% compared with Fable 5.
These figures highlight the distinction between benchmark performance and production economics. A model does not necessarily need to achieve the highest benchmark score to be useful if it can complete a comparable task at a fraction of the operating cost.
Computer Use Efficiency #
Computer-use workloads show a similar pattern.
In the OSWorld 2.0 offline evaluation, GPT-6 Sol at maximum reasoning effort achieved a result close to Claude Opus 5 at medium effort:
- GPT-6 Sol: 60.5%
- Claude Opus 5: 60.3%
OpenAI reports that GPT-6 Sol’s per-task cost was approximately 80% lower in this comparison.
GPT-6 Luna at maximum reasoning effort also surpassed GPT-5.6 Sol at medium effort while operating at approximately one-tenth of the cost.
Prompt Caching for Long-Horizon Agents #
OpenAI has also made changes to caching for long-running agent workloads.
GPT-6 introduces improved Prompt Caching behavior intended to increase default cache hit rates. Agents can reuse more previously processed context, potentially reducing latency and inference expenditure for workflows that repeatedly access the same code, documents, or conversation history.
Cached input tokens also receive a 90% discount.
For agentic systems, this can be as important as reducing headline input and output token prices because long-running tasks can repeatedly consume large amounts of contextual information.
The broader implication is that AI economics increasingly depend on the complete inference stack rather than model token prices alone.
🤖 Anthropic Introduces Claude Opus 5.5 #
Anthropic’s response is Claude Opus 5.5, the first model in the Claude 5.5 series.
Anthropic positions Opus 5.5 around complex coding, knowledge work, agentic workflows, and computer use. Compared with Opus 5, the company says Opus 5.5 reduces typical task costs by approximately 40% while increasing output speed by more than 30%.
Anthropic’s announcement:
https://www.anthropic.com/claude-opus-5-5/
The company also presents Opus 5.5 as a model capable of reaching or exceeding the performance of Claude Fable 5.1 on several evaluations while operating at a much lower price.
Benchmark Performance #
Anthropic’s published comparison shows Opus 5.5 ahead of Fable 5.1 on several listed evaluations.
On Terminal-Bench 4.0, Opus 5.5 scored 66.4%, compared with 55.8% for Fable 5.1.
On FrontierCode v1.1 Main, Opus 5.5 scored 54.4%, compared with 50.3% for Fable 5.1.
On GDPval-AA v2.1, which evaluates real-world knowledge work, Opus 5.5 reached 1846 compared with 1735 for Fable 5.1.
These evaluations emphasize the areas Anthropic is targeting with Opus 5.5: agentic coding, knowledge work, and computer interaction.
However, benchmark comparisons should be interpreted in context. Different evaluations use different harnesses, tools, datasets, and reasoning configurations. Anthropic’s own published comparisons also show GPT-6 Astra ahead of Opus 5.5 on selected evaluations such as AutomationBench and Terminal-Bench-Science.
Therefore, individual benchmark results do not establish a universal performance hierarchy between the models.
Lower Pricing for Opus 5.5 #
Claude Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens.
Claude Fable 5.1 is priced at $10 per million input tokens and $50 per million output tokens, making its listed input and output rates 2.5 times higher.
Both models provide a 1-million-token context window and up to 128K tokens of standard output capacity.
The pricing structure reinforces Anthropic’s focus on cost per completed task rather than token price alone.
🧠 Adaptive Thinking Changes the Economics of Reasoning #
One of the more significant changes in Opus 5.5 is its Adaptive Thinking system.
Adaptive Thinking is enabled by default, while developers can adjust reasoning effort. The default setting is medium, whereas Fable 5.1 defaults to high effort.
Anthropic’s FrontierCode v1.1 Main results illustrate why this distinction matters.
At medium effort, Opus 5.5 achieved approximately 54.6% with a per-task cost of around $0.80. At maximum effort, the cost increased to approximately $6.19, while the score fell slightly to 54.4%.
In this particular evaluation, increasing reasoning expenditure therefore did not improve the measured result.
One explanation is that FrontierCode penalizes modifications outside the requested task scope. Additional reasoning can cause a model to perform extra changes that may be technically useful but negatively affect the benchmark score.
This illustrates an important limitation of test-time compute: more reasoning does not automatically translate into better task outcomes.
Third-party testing from CodeRabbit reported a similar pattern in multi-step code review workflows, where Opus 5.5 at medium effort could match or exceed Opus 5 at high effort while consuming approximately half as many tokens.
💰 Why Task Costs Can Fall Faster Than Token Prices #
Anthropic’s pricing structure also demonstrates why token prices alone do not determine the actual cost of an AI workflow.
Compared with Opus 5, standard input and output pricing for Opus 5.5 fell by approximately 20%. However, Cache Read pricing dropped from $0.50 to $0.20 per million tokens, representing a 60% reduction.
This is particularly important for long-context workloads such as Claude Code, where the same codebase, conversation history, or task context may need to be processed repeatedly.
Cache savings accumulate across these repeated operations.
When lower cache costs are combined with Opus 5.5 completing tasks using fewer tokens and fewer reasoning steps, Anthropic reports that typical task costs fall by approximately 40% under default settings.
This distinction matters for production deployments. A model that costs 20% less per token does not necessarily reduce the final cost of completing a workflow by only 20%.
The actual calculation depends on:
- Number of tokens consumed
- Number of inference steps
- Cache hit rates
- Reasoning effort
- Number of retries
- Tool calls
- Task completion rate
- Human intervention requirements
For production AI systems, total workflow cost is therefore a more useful operational metric than headline token pricing alone.
📊 AI Competition Is Shifting Toward Cost per Completed Task #
Taken together, the OpenAI and Anthropic releases highlight a broader change in the economics of generative AI.
The central question is increasingly not how intelligent a model can become, but how much it costs to reliably complete a useful piece of work.
OpenAI discussed this concept in its earlier “Building Abundant Intelligence” article, arguing that customers ultimately purchase completed tasks rather than raw tokens.
https://openai.com/index/building-abundant-intelligence
The total cost of obtaining a successful result can include inference, retries, human oversight, latency, and the cost of correcting model errors.
A model with a higher token price can therefore be cheaper in practice if it completes a task using fewer tokens, requires fewer retries, or produces more reliable results.
Enterprise Pricing Pressure #
Anthropic has also received evidence of enterprise sensitivity to model pricing.
Following the release of Fable 5, Ramp’s enterprise AI spending tracker indicated that Anthropic’s most expensive and capable model represented approximately 6% of enterprise token usage.
That pattern can be interpreted as a signal that enterprises do not automatically move all workloads to the most capable model when the price difference is substantial.
Anthropic subsequently introduced pricing and caching changes with Fable 5.1. At launch, the company cited customer feedback concerning pricing, data retention, and safety. Although standard input and output list prices remained unchanged, Cache Read pricing was reduced by 75%, with Anthropic reporting typical workload savings of approximately 25% and reductions of up to 45% for highly agentic tasks.
Anthropic’s developer guidance also recommended starting most tasks with Opus 5 and reserving Fable 5.1 for situations where high-effort Opus was insufficient.
This reflects a broader deployment pattern: flagship models increasingly function as specialized resources for difficult, long-horizon workloads rather than as the default model for every request.
🔍 From Maximum Intelligence to Efficient Intelligence #
The latest releases from OpenAI and Anthropic suggest that frontier-model competition is entering a new phase.
Model capability remains important, but production adoption depends on a larger set of variables:
- Task completion quality
- Cost per successful task
- Inference latency
- Token efficiency
- Cache utilization
- Reasoning efficiency
- Reliability and retry rates
- Tool-use and agent performance
This changes how developers should evaluate AI models.
A model that achieves a slightly higher benchmark score may not be the most economical choice for a production workload if another model reaches the required quality level at substantially lower cost.
Likewise, a model with a lower token price can still be expensive if it requires more reasoning steps, consumes more context, or needs frequent retries.
The practical unit of comparison is increasingly the completed workflow rather than the individual token.
🌐 The Next Phase of LLM Competition #
OpenAI and Anthropic are approaching the same problem from somewhat different directions, but their latest releases converge on an important theme: frontier AI must become substantially cheaper to use if it is going to expand across large-scale production workloads.
GPT-6 Sol and GPT-6 Luna emphasize lower API prices, efficient inference, caching, coding performance, and agent workloads.
Claude Opus 5.5 emphasizes adaptive reasoning, lower effective task costs, faster output, cache efficiency, and strong performance in coding and knowledge work.
The competition therefore extends beyond benchmark leadership. It increasingly concerns how much useful work can be produced for every dollar of inference spending.
This also explains why lower-tier models are becoming strategically important. If a production workload does not require the maximum capability available from a frontier model, a faster and substantially cheaper model can support far greater throughput.
For developers, the key metric is increasingly straightforward: how much does it cost to reliably finish the job?
As AI systems move from experimentation into continuous production use, that metric may become as important as raw benchmark performance in determining which models developers deploy at scale.