OpenAI has officially launched GPT-5.6, and it's not just another incremental update. The new model family — led by the flagship Sol, with Terra for balanced everyday work and Luna for cost-efficient tasks — represents a genuine step change in what AI can do for businesses. From coding and knowledge work to cybersecurity and scientific research, GPT-5.6 sets new state-of-the-art benchmarks while simultaneously driving down cost and latency. Here's what you need to know.

Three Tiers, One Generation

GPT-5.6 introduces a naming convention worth understanding. The number (5.6) identifies the generation. Sol, Terra, and Luna are durable capability tiers that can advance on their own cadence — meaning OpenAI can upgrade Terra or Luna without necessarily moving to GPT-5.7.

  • Sol — The flagship. State-of-the-art results across every benchmark that matters. Priced at $5/$30 per million tokens (input/output).
  • Terra — Balanced performance, competitive with the previous GPT-5.5 generation at a lower cost. $2.50/$15 per million tokens.
  • Luna — Fast and affordable, yet still outperforming many previous frontier models. $1/$6 per million tokens.

The pricing spread matters. Luna outperforms GPT-5.5's peak on several benchmarks at less than half the cost. For high-volume applications — customer support, content classification, data extraction — that's a dramatic reduction in the economics of AI deployment.

Intelligence Per Dollar: The Real Story

Raw benchmark numbers are interesting. Performance per dollar is what actually drives business decisions. And this is where GPT-5.6 pulls ahead decisively.

On Agents' Last Exam — an evaluation of long-running professional workflows across 55 fields — GPT-5.6 Sol scores 53.6, beating Claude Fable 5 by 13.1 points. Even at medium reasoning, it outperforms Fable 5 by 11.4 points at roughly one-quarter the estimated cost. Terra and Luna both outperform Fable 5 at around one-sixteenth the cost.

On the Artificial Analysis Intelligence Index, a broad measure spanning agentic work, coding, scientific reasoning, and general capabilities, GPT-5.6 Sol with max reasoning comes within one point of Fable 5 — while completing tasks in 61% less time at roughly half the cost.

"The result is stronger performance per dollar: more successful work for the same spend, or comparable results at a lower total cost."

Coding: A New State of the Art

If there's one area where GPT-5.6 delivers an unambiguous win, it's code. On the Artificial Analysis Coding Agent Index, GPT-5.6 Sol with max reasoning scores 80 — 2.8 points above Claude Fable 5 — while using less than half the output tokens, taking less than half the time, and costing about one-third less.

The advantage extends across the family. Terra performs just above Fable 5. Luna outperforms Claude Opus 4.8. Each achieves this in roughly one-third of the time, with about half as many output tokens, at approximately one-quarter the estimated cost.

GPT-5.6 also sets new records on Terminal-Bench 2.1 (88.8% for Sol, 91.9% for Sol with ultra) and DeepSWE (72.7%), which test complex command-line workflows and long-horizon engineering in real codebases. This isn't just about writing functions — it's about navigating, modifying, and shipping within large existing projects.

Programmatic Tool Calling

One of the most practically useful new features is Programmatic Tool Calling. Instead of passing every tool response back through the model for the next decision, GPT-5.6 can write and run lightweight programs in-memory that coordinate tools, filter intermediate data, monitor progress, and choose the next action autonomously.

For developers, this means fewer tokens, fewer model round trips, and less hand-holding. The model can process a large dataset, extract what matters, discard the rest, and adapt its approach — all within a single request. And it's Zero Data Retention compatible, which matters for enterprise security teams.

Ultra: Multi-Agent at the Model Level

The headline new capability is ultra — GPT-5.6's highest-capability setting. By default, ultra coordinates four agents in parallel across separate workstreams. For the most demanding tasks, developers can scale this to 16 agents via the multi-agent beta in the Responses API.

The benchmarks tell the story. On BrowseComp, SEC-Bench Pro, and Terminal-Bench 2.1, adding parallel agents shifts the score-latency frontier upward and to the left — meaning stronger results in less wall-clock time. This is the first time a major model provider has built multi-agent orchestration directly into the model layer rather than leaving it to application code.

For businesses, this means complex tasks — multi-source research, large-scale code refactoring, comprehensive document analysis — can be completed faster without the engineering overhead of building and debugging your own agent coordination framework.

Knowledge Work and Computer Use

GPT-5.6 delivers what OpenAI describes as a "step change in design judgment." With only high-level direction, it can create polished, functional interfaces — and critically, it can inspect and refine the rendered result, catching visual and functional issues before handing work back.

This shows up in the benchmarks:

  • BrowseComp: 92.2% (new state of the art) — agentic browsing tasks that require navigating, reading, and synthesising information across the web.
  • OSWorld 2.0: 62.6% — computer-use tasks that require operating real software interfaces. GPT-5.6 surpasses Opus 4.8 while using 85% fewer output tokens.
  • BenchCAD: 70.6% (83.4% with python tool) — design and engineering tasks.

For professional workflows, GPT-5.6 can ingest messy context from documents, Slack, Notion, Microsoft 365, and Google Drive, then convert it into expert-level, shareable artifacts — presentations, spreadsheets, reports — with strong layouts, consistent typography, and accurate formatting. It can infer a deck's design system from a reference template and apply those conventions to new material.

Cybersecurity: Dual-Use Done Right

GPT-5.6 is OpenAI's strongest cybersecurity model yet — and they've clearly thought hard about the dual-use problem.

On the offensive side:

  • ExploitBench2: 73.5% (vs GPT-5.5's 47.9%) — progressing from reaching vulnerable code to arbitrary code execution.
  • ExploitGym3: 24.9% under 2-hour cap (up from 15.1%), reaching 33.7% with 6 hours — turning real-world vulnerabilities into working exploits.
  • SEC-Bench Pro: 71.2% (vs GPT-5.5's 45.8%) — proof-of-concept generation on complex software.

On the defensive side, GPT-5.6 supports secure code review, patching, threat modelling, and blue teaming. OpenAI's Trusted Access for Cyber program gives verified security professionals deeper access to defensive capabilities — vulnerability triage, malware analysis, detection engineering, and patch validation — with safeguards calibrated to the trust level of the user and environment.

Importantly, OpenAI's testing suggests GPT-5.6 is better at finding and fixing vulnerabilities than at autonomously carrying out end-to-end attacks against hardened targets. That asymmetry favours defenders — which is how it should be.

Science and Research Acceleration

GPT-5.6 Sol shows broad gains across scientific research, with Pareto improvements over GPT-5.5 on real-world biology, life science research workflows, and chemistry. On GeneBench Pro (genomics and quantitative biology), it reaches stronger results with fewer tokens and less time.

Perhaps most strikingly, OpenAI reports that internal researchers using GPT-5.6 more than doubled their daily output tokens compared to peak GPT-5.5 usage. Over the past six months, internal research compute devoted to coding inference grew 100-fold, and agentic token usage increased 22-fold. On their internal recursive self-improvement (RSI) evaluation bundle, GPT-5.6 Sol shows a 16.2-point improvement over GPT-5.5.

Safety: Layered, Adaptive, Reasoning-Based

GPT-5.6 launches with what OpenAI calls their "most robust safety system to date." The approach is layered and worth understanding for any organisation considering deployment:

  1. Trained protections — Built into the model itself during training.
  2. Reasoning monitor — Instead of relying solely on classifier flags (which are slow to update and prone to false positives), a reasoning model reviews the conversation in real-time to assess potential for harm. This can be updated rapidly without retraining classifiers.
  3. Real-time checks and monitoring — Continuous evaluation during use.
  4. Account-level enforcement — Access calibrated to trust and risk, with hardware-backed passkeys required for the most sensitive capabilities by September 2026.

The models do not cross the Critical threshold in either biology or cybersecurity. OpenAI also notes that their GPT-5.6 Sol cyber safeguards block roughly ten times more potentially harmful activity compared to previous models — starting conservatively and improving based on real-world use.

Availability and Access

GPT-5.6 is rolling out globally across ChatGPT, Codex, and the OpenAI API, with full availability within 24 hours of launch.

  • ChatGPT: Plus, Pro, Business, and Enterprise users get GPT-5.6 Sol at medium and higher effort settings. Pro and Enterprise can select Sol Pro for maximum quality.
  • ChatGPT Work and Codex: Free and Go users access Terra. Paid tiers can choose among Sol, Terra, and Luna, with effort-level controls. The max setting is available to all users; ultra is available to Pro and Enterprise (ChatGPT Work) or Plus and higher (Codex).
  • API: All three tiers are available. Programmatic Tool Calling is ZDR-compatible. Multi-agent is in beta.

Prompt caching has been improved with explicit cache breakpoints, a 30-minute minimum cache life, and cache writes billed at 1.25x the uncached input rate (reads still get the 90% discount).

What This Means for Your Business

GPT-5.6 isn't just a better model — it's a shift in the economics and capabilities of AI for business. Here's our take on what matters most:

  • The cost floor is dropping fast. Luna at $1/$6 per million tokens, outperforming last-generation flagships, means use cases that weren't economically viable six months ago — high-volume content classification, real-time data extraction, conversational AI at scale — now are.
  • Coding productivity is entering a new phase. With Terminal-Bench scores above 88% and Programmatic Tool Calling reducing round trips, development teams can delegate significantly more complex tasks. The multi-agent ultra mode means large refactoring projects or multi-file feature work can be parallelised at the model level.
  • Knowledge work is becoming automatable end-to-end. The combination of BrowseComp, OSWorld, and design judgment means GPT-5.6 can take messy inputs — Slack threads, scattered documents, half-formed briefs — and produce polished, shareable deliverables. That's not a chatbot. That's a junior analyst.
  • Security teams get a powerful new tool. The defensive capabilities — secure code review, vulnerability triage, patch validation — are genuinely useful, and the Trusted Access program ensures they're available to verified professionals without giving attackers the same advantage.
  • Multi-agent is no longer experimental. With ultra mode built into the model layer, businesses can leverage parallel agent coordination without building and maintaining their own orchestration infrastructure.

The Bottom Line

GPT-5.6 represents the most significant efficiency leap in the current generation of frontier models. It's not just smarter — it's dramatically cheaper per unit of work, faster on complex tasks, and more capable of operating autonomously across tools and environments. For businesses evaluating AI investments, the performance-per-dollar improvements across all three tiers make this the moment to reassess what's possible.

The question isn't whether GPT-5.6 can do more. It's whether your organisation is ready to take advantage of what it offers.