GPT-5.6: what OpenAI announced

OpenAI’s GPT‑5.6 announcement is dense with benchmark numbers and efficiency claims. The company positions Sol as the flagship model, with Terra and Luna filling in for balanced and cost‑efficient tiers. The through line is not just higher raw scores. It is capability per token and time to result across agentic workflows, coding, browsing, science, and cyber tasks, with repeated emphasis on lower cost and latency for comparable outcomes. I read the page with one question in mind: is Fable 5 beaten?

OpenAI’s own comparisons are the only basis here, and even they frame “better” in different ways depending on the test. On some evaluations the margin is large. On others the story is parity or near‑parity on score while claiming big gains on speed and cost. In a few areas, comparison with Claude Fable 5 is not shown because it refused to answer those evals. That nuance matters if the question is whether Fable 5 is uniformly surpassed.

Model family and reasoning settings

OpenAI introduces a three‑model family. Sol is the flagship, Terra is described as balanced, and Luna is the cost‑efficient option. Alongside the models, there are new reasoning and performance settings. The default aims for efficiency. There are xhigh and max settings that allow more time and compute for harder problems, and an ultra setting that coordinates multiple agents. Ultra defaults to four agents working together. The page shows that these settings are not just knobs for more tokens. They are meant to shift the quality, speed, and cost tradeoffs in structured ways.

OpenAI’s charts use those settings to anchor comparisons. For multi‑agent scenarios specifically, they show ultra pushing the score‑latency frontier up and to the left on tasks like BrowseComp, SEC‑Bench Pro, and Terminal‑Bench 2.1. In the API, a multi‑agent beta in the Responses API is available so developers can build similar coordinated experiences.

Headline benchmarks and the Fable 5 question

On Agents’ Last Exam, which OpenAI describes as long‑running professional workflows across 55 fields, GPT‑5.6 Sol scores 53.6. OpenAI reports that Sol eclipses Claude Fable 5 with adaptive reasoning by 13.1 points. They also argue this gap holds in a cost‑sensitive setting: at medium reasoning, Sol beats Fable 5 by 11.4 points at roughly one‑quarter the estimated cost. Terra and Luna are reported to outperform Fable 5 at around one‑sixteenth the cost. If your definition of “beaten” includes both score and cost, this is one of the places where the claim looks strongest based on OpenAI’s reporting.

On the Artificial Analysis Intelligence Index, the picture is closer. With max reasoning, OpenAI says Sol comes within one point of Fable 5 while completing tasks in 61 percent less time at roughly half the estimated cost. Here the comparative story is not a clear score win. It is a near‑tie that favors GPT‑5.6 on speed and cost.

Coding and agentic engineering

For code‑heavy work, OpenAI highlights the Artificial Analysis Coding Agent Index. GPT‑5.6 Sol with max reasoning scores 80, which the page says is 2.8 points above Fable 5. The efficiency framing is explicit: Sol uses less than 50 percent of the output tokens, takes less than 50 percent of the time, and costs about one third of the estimated cost. OpenAI also claims new state of the art on Terminal‑Bench 2.1 and DeepSWE, two agentic and code‑base‑focused tests.

The announcement pairs these scores with a product point. Programmatic Tool Calling in the Responses API is described as a way to filter intermediate tool data, retain only what matters, and adapt workflow to reduce round trips and tokens. Read together, the benchmarks and the API feature are meant to show not just raw model IQ but a path to building leaner multi‑tool agents.

Knowledge work, browsing, and design

On browsing and generalist tasks, OpenAI reports that GPT‑5.6 Sol achieves a 92.2 percent score on BrowseComp, which they call state of the art. On OSWorld 2.0, Sol scores 62.6 percent and surpasses Opus 4.8 while using 85 percent fewer output tokens. These are framed as examples of a better score‑per‑token profile rather than only absolute accuracy.

OpenAI also claims stronger design judgment and better conversion of messy context from tools like Slack, Notion, Microsoft 365, and Google Drive into shareable artifacts. For deck generation specifically, they point to better fidelity to templates and Slide Masters. These are qualitative assertions on the page that sit alongside the quantitative browsing results to position GPT‑5.6 as a more useful partner for end‑to‑end knowledge work.

Cybersecurity capabilities and access controls

Security is an area where the page shows large jumps over GPT‑5.5. On ExploitBench 2, GPT‑5.6 scores 73.5 percent versus GPT‑5.5’s 47.9 percent at comparable output‑token budgets. On ExploitGym 3, the pass rate improves from GPT‑5.5’s 15.1 percent to 24.9 percent under a two‑hour cap, and with six hours reaches 33.7 percent. On SEC‑Bench Pro, GPT‑5.6 scores 71.2 percent versus GPT‑5.5 at 45.8 percent. The defensive framing is explicit too. OpenAI states GPT‑5.6 supports secure code review, patching, threat modeling, and blue teaming, with more defensive capability available to qualified users through Daybreak’s Trusted Access for Cyber program.

The access model for these capabilities is tightened. Individuals can verify identity and request access, and organizations can apply. Advanced Account Security with hardware‑backed passkeys will be required by September 1 to retain access to most cyber‑capable frontier models, and OpenAI names Yubico as a partner for preferred pricing. The company also states that it is taking steps to restrict access to high‑risk entities and jurisdictions. For a backend engineer focused on security, the combination of stronger offensive and defensive capability with stricter access control is the operational story: more powerful tools, more friction to get to them, and a clear expectation of identity assurance.

Life sciences and chemistry

OpenAI reports Pareto improvements over GPT‑5.5 across real‑world biology, life‑science workflows, and chemistry. GeneBench Pro, LifeSciBench, and MedChemBench are cited as evidence on the page. A notable caveat appears in the comparison set. The page says Claude Fable 5 was not included in advanced biology evaluations because it “does not answer advanced biology questions and refuses the majority of questions in this eval.” If you read “beaten” as universal superiority, this is one of the places where the comparison is not directly available.

Efficiency, latency, and internal adoption

The efficiency narrative repeats across the announcement. OpenAI includes multiple customer and partner testimonials on token efficiency, latency, and quality across domains like coding, decks, frontend tasks, and financial research. The specifics are on their page, but the overall pattern is consistent with the benchmark framing: similar or better outcomes at lower token and time budgets.

OpenAI also shares internal metrics to show adoption. They report that average daily output tokens per active researcher are more than twice the highest level seen for GPT‑5.5. The share of research compute for internal coding inference grew 100‑fold. Internal agentic token usage increased by roughly 22‑fold. The company is careful to add that these adoption metrics “do not measure research progress on their own.” They are inputs and signals, not direct evidence of scientific breakthroughs.

Safety process and evaluation scope

The launch followed an evaluation period that, according to OpenAI, included human red teaming, large‑scale automated testing, work with expert organizations, and trusted partners. The page describes a layered approach to safety with protections trained into the model, real‑time checks, monitoring, and access calibrated to trust and risk. Those details give context to the access controls around cyber capabilities and the decision to limit some features to verified individuals and organizations.

The scope and framing of the evaluations also matter for comparisons. Where OpenAI reports Claude Fable 5 comparisons, sometimes the margins are large, and sometimes they show near parity with an efficiency edge. In life sciences, Fable 5 is omitted for refusal reasons. That is a material caveat when reading across domains.

So, is Fable 5 beaten?

On OpenAI’s own reporting, GPT‑5.6 Sol looks like a substantial step for agentic, coding, browsing, and cyber tasks. On Agents’ Last Exam, Sol’s 53.6 and the 13.1‑point gap over Fable 5 is a clear win, with additional claims of better cost efficiency at medium reasoning and even Terra and Luna outperforming at a fraction of the cost. On the Artificial Analysis Intelligence Index, Sol with max reasoning comes within one point of Fable 5, and OpenAI emphasizes completing tasks 61 percent faster at roughly half the estimated cost. Coding shows a similar pattern: Sol at 80 on the Coding Agent Index, 2.8 points above, with less than half the tokens and time at about one third the cost.

If “beaten” means score plus cost and latency across many practical workflows, OpenAI’s page gives them a strong case in several categories. If “beaten” means undisputed raw‑score superiority everywhere, the answer is more qualified. Some tests are near ties with an efficiency edge, and in advanced biology OpenAI does not include Fable 5 because it refused most questions in that evaluation. The broader takeaway I drew is that GPT‑5.6 is being positioned as the more capable‑per‑token and faster path for agentic systems, coding, and security work. Whether that settles the Fable 5 question depends on which benchmark you care about and how much weight you put on cost and latency alongside score.

References

• GPT‑5.6 announcement