GPT-5.6 Slashes Agent Costs

OpenAI is pushing the envelope on AI agent economics with the release of its GPT-5.6 model family. The company claims these new models deliver frontier-level agent performance at a dramatically lower cost, a move that could significantly accelerate the adoption of sophisticated AI assistants across industries. The announcement details how startups are already leveraging smarter model selection and new API controls to build faster, more capable agents.

The core promise of GPT-5.6 is a better out-of-the-box experience. Building on advancements from previous generations, GPT-5.6 aims to tackle longer tasks with fewer tokens. This translates to stronger agent performance and lower costs with minimal changes to existing infrastructure. OpenAI highlights improved accuracy at lower reasoning efforts, citing examples where GPT-5.6 at a "low" reasoning setting outperformed GPT-5.5 at a "high" setting on the Agents' Last Exam benchmark. Startups involved in early testing have reported substantial cost savings across various workflows by reducing the default reasoning effort.

Model Selection for Efficiency

Traditionally, achieving peak performance for complex, long-horizon tasks meant opting for the most powerful, and expensive, flagship models. This often involved using them for all parts of a task, even simpler ones. GPT-5.6 shifts this paradigm. The new Luna and Terra models within the 5.6 family can match or exceed the performance of older models like GPT-5.4 and 5.5, but at a fraction of the cost. This makes sophisticated capabilities like document understanding and complex browser interactions economically viable for a wider range of applications.

Companies like Hypha are seeing dramatic improvements. Serhii Shchoholiev, Engineering Lead at Hypha, noted that Luna retains 98% of GPT-5.5's extraction accuracy at one-eighteenth the cost. Similarly, Gregor Zunic, Co-Founder at Browser Use, reported that Luna completed 78% of challenging browser tasks for approximately $14, compared to roughly $235 for the previous state-of-the-art model to achieve 80% completion. PlayerZero's Founder and CEO, Animesh Koratana, adopted Luna for high-throughput code retrieval and decision modeling, resulting in a 64% cost reduction, a 90% cut in response time, and a five-point F1 score improvement.

Even benchmarks designed to stress AI models show significant gains. On BrowseComp, a search-based benchmark for obscure facts, GPT-5.6 Luna (Extra High) achieved 84.04% performance at a cost of $1.33, a stark contrast to GPT-5.5 (Extra High) which scored 84.36% for $33.27. OpenAI has since further reduced prices.

Evolving the Responses API

Beyond the models themselves, OpenAI has introduced new primitives to the Responses API designed to build more efficient agents. These architectural interventions focus on three key areas: reusing previous work through persistent reasoning and conversation compaction, enabling parallel decomposition with native multi-agent orchestration, and moving deterministic work into code via programmatic tool calling.

Retained reasoning and compaction allow agents to maintain coherence over longer tasks without losing context or needing to recompute prior steps. Multi-agent orchestration facilitates parallel workstreams for complex tasks, speeding up completion. Programmatic tool calling, which enables GPT-5.6 to write JavaScript for orchestrating tools, filtering data, and processing outputs outside the model's context window, is particularly impactful. This reserves the model's expensive token usage for judgment and reasoning, reducing cost and latency.

On the ARC-AGI-3 benchmark, enabling retained reasoning and compaction for GPT-5.6 Sol boosted its score from 13.3% to 38.3%, while using approximately six times fewer output tokens. This demonstrates how architectural improvements can dramatically amplify model performance without changes to the core model itself.

Programmatic Tool Calling and Multi-Agent Workflows

The distinction between judgment-based tasks and data-intensive work is central to efficient agent design. Programmatic Tool Calling allows agents to handle the latter, like retrieving, filtering, and combining data from numerous sources, using external code. Alex Wang from Rogo noted that for financial research, this capability matched their rubric quality while reducing input tokens by 21%, enabling agents to perform actual research rather than just discussing it.

For complex, parallelizable tasks, OpenAI's native multi-agent orchestration, now available in the Responses API, allows for distributing actions and reasoning across multiple agents. A primary agent coordinates subagents, which work in parallel and return their findings for final synthesis. E Chi, Founder of Quadrillion, found GPT-5.6 Sol to be an excellent orchestrator for open-ended research problems, outperforming GPT-5.5 and many other tested models. Jon Bell, Co-founder and CPO at Obvious, praised GPT-5.6 as the best orchestrator seen from OpenAI, successfully managing six complex tasks simultaneously without quality degradation.

Prompt Caching Enhancements

To further reduce latency and cost, OpenAI has extended the prompt cache Time To Live (TTL) to a minimum of 30 minutes across the model family. Deterministic cache breakpoints within the context window have also been introduced. Lorenzo Gentile, an AI Engineer at Ploy, reported a 28% reduction in uncached input by adding cache breakpoints and workspace-specific keys to a large prompt. The extended cache window allows agents to reuse context across runs, avoiding redundant computation.

The economic shift is clear: use cases that once demanded expensive frontier models at every turn can now achieve comparable or superior results by combining smaller, cost-optimized models, fine-tuning reasoning efforts, and employing these new architectural choices. This development promises to democratize access to sophisticated AI agent capabilities, making them a practical reality for a much wider array of startups and businesses.