You Can Keep the Benchmarks. I’ll Take the Test Drive

OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 represent a substantial advance in AI, demonstrating improved understanding of complex assignments and agentic capabilities that reduce the need for human guidance. The models score 59 and 60 respectively on Artificial Analysis's Intelligence Index, confirming their leading-edge performance.

TL;DR — Key Takeaways Benchmarks matter—but real-world performance matters more. The latest frontier AI models are proving their value by handling complex, multi-step work rather than simply scoring well on tests. AI is becoming agentic. New models can plan, use tools, recover from mistakes, and carry projects forward with far less human guidance than previous generations. The AI race is bigger than OpenAI vs. Anthropic. Google, xAI, Meta, Microsoft, Mistral, DeepSeek, Alibaba, Zhipu, and others are rapidly advancing capabilities, creating fierce global competition. Smarter AI also raises bigger security risks. As models gain access to tools, code repositories, cloud environments, and business systems, governance, identity controls, and approval boundaries become critical. I have never been one to judge technology by looking at a chart. Benchmarks are useful, especially when the technology is as difficult to evaluate as artificial intelligence. They provide common tests, measurable results and some basis for comparing one model with another. The problem is that technology companies rarely publish charts in which their own product finishes second. My test has always been more practical. Give me the technology and let me use it. Give it real work rather than a carefully constructed demonstration. Does it understand what I am trying to accomplish? Can it sustain quality across a long project? Does it require constant correction? Most importantly, does it remain a tool I have to operate one prompt at a time, or can it begin taking some responsibility for completing the work? By that admittedly subjective measure, the newest generation of frontier models represents a substantial advance. The two models I have used enough to judge are OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5. I have not fed them identical prompts under laboratory conditions, timed their responses or invented a scoring system that assigns points for every strength and weakness. I have put both of them to work on the research, writing, editing, analysis and production assignments that occupy my days. My conclusion is not especially close. Both are significant steps up from what came before. The improvement is visible in the quality of their output, but that may be the least important part of it. Previous frontier models could produce excellent work. They could also produce an excellent first page followed by four pages of repetition, lose the original thesis halfway through an assignment or require so much prompting and correction that the user remained the only real project manager in the room. Sol and Fable are better at understanding the whole assignment. They maintain intent across longer bodies of work, recognize relationships among source materials and make more useful judgments about what belongs, what does not and what still needs to be done. They require less steering and are more capable of recovering when an initial approach does not work. Most significantly, they arrive with agentic capabilities that move them beyond simply answering a prompt. They can inspect available materials, develop a plan, call tools, search for missing information, create artifacts, check their work and continue through a multistep process. They are beginning to participate in the work rather than merely comment on it. The Data Catches Up With the Experience Because my impressions are subjective, it is worth asking whether the objective evidence supports them. So far, it does. Artificial Analysis, one of the more useful independent organizations evaluating AI models, currently gives Claude Fable 5 a score of 60 on its Intelligence Index. GPT-5.6 Sol is immediately behind it at 59. The one-point gap is far less meaningful than the shared result: Both models occupy the leading edge of the current field. More interestingly, Artificial Analysis estimates https://artificialanalysis.ai/articles/gpt-5-6-has-landed that Sol reaches roughly the same intelligence tier at approximately one-third the evaluated cost of Fable. Sol also leads the Artificial Analysis Coding Agent Index with a score of 80. That result is particularly relevant because coding-agent evaluations measure something closer to sustained work than traditional question-and-answer benchmarks. The model must understand a task, navigate an environment, make changes and determine whether those changes worked. Anthropic has published similarly strong results for Fable 5. The company says the model became the first to exceed 90% on one of its benchmarks for complex, long-running analytical tasks, a ten-point improvement over its previous Opus model. That is vendor-supplied evidence and should be treated accordingly, but it is consistent with my experience. Fable is particularly good at holding onto the logic and texture of a large assignment instead of reducing it to a series of disconnected prompts. The broader field shows that this is less a two-horse race than a rapidly advancing pack, with the highest scores concentrated among the newest releases. Figure 1. Selected frontier models ranked using the Artificial Analysis Intelligence Index v4.1. Higher is better. The chart confirms the advance, but it also shows why declaring one permanent winner would be foolish. The field is compressed, the models are optimized for different purposes and the order is changing with nearly every release. A score of 60 does not prove that Fable will be better than Sol for your work, any more than a score of 59 proves that Sol will be the better value in every situation. Benchmarks measure performance under defined conditions. They do not measure the fit between a user and a model, how well a model adapts to a working style or how much confidence someone develops in assigning it progressively more difficult work. Still, when independent measurements and practical experience point in the same direction, the result is hard to dismiss. This is Bigger Than OpenAI and Anthropic The current release cycle makes clear that this is bigger than another OpenAI-versus-Anthropic contest. xAI has released Grok 4.5 https://x.ai/news/grok-4-5 , positioning it specifically for coding, agentic tasks and knowledge work. Its Artificial Analysis Intelligence Index score of 54 places it below Fable and Sol but firmly within the frontier field. I have not used Grok 4.5 enough to offer an informed personal judgment. I intend to, particularly because xAI is emphasizing the same transition from answering questions to carrying out longer and more complicated assignments. Google has released Gemini 3.5 Flash https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/ , with the more powerful Gemini 3.5 Pro expected to follow. Flash currently trails the premium leaders on raw intelligence measures, but raw intelligence is only one part of the equation. Google is emphasizing speed, economics, coding and long-horizon agentic work. A model that delivers enough intelligence much faster and at a substantially lower cost may be more valuable for many production workloads than the model that tops the leaderboard. Meta has reentered the conversation with Muse Spark 1.1 https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/ . It scores 51 on the Artificial Analysis Intelligence Index, an eight-point gain over the original Muse Spark, while combining that performance with high output speed and aggressive pricing. Meta appears to have made meaningful progress in tool use, computer use, coding and multimodal reasoning. Once again, the advances point toward agency rather than better conversation alone. Microsoft is harder to place in a simple model horse race. Its most important work may not be a general-purpose model directly comparable with Sol or Fable. Microsoft is developing specialized models, computer-use agents, multi-model systems and an expanding agentic layer across its enterprise products. The Fara 1.5 family, for example, is focused on computer use rather than winning a general intelligence leaderboard. I have not evaluated Microsoft’s newest agentic systems sufficiently to judge them, but its distribution through Windows, Azure, GitHub and Microsoft 365 could ultimately matter more than whether one Microsoft model wins a laboratory test. Then there is China, which no longer belongs in an obligatory paragraph at the end of an article about frontier AI. Zhipu’s GLM-5.2 currently scores 51 on the Artificial Analysis Intelligence Index, matching Meta’s Muse Spark 1.1 and leading the open-weight field. MiniMax-M3 and DeepSeek V4 Pro https://api-docs.deepseek.com/news/news260424 are within striking distance, with DeepSeek combining competitive reasoning with the low costs and open availability that made its earlier models so disruptive. Alibaba’s Qwen 3.6 and Qwen 3.7 https://qwen.ai/blog?id=qwen3.7 families are increasingly designed around real-world agents, coding, tool use and operation within simulated environments. These models do not need to finish first on every Western benchmark to reshape the market. Their combination of capability, openness and economics limits how much of a premium the closed-model leaders can charge. They also ensure that advanced model development will not remain confined to a handful of American companies. Mistral occupies another important position. Mistral Large 3 https://mistral.ai/news/mistral-3/ remains the company’s broad flagship, but Mistral’s strategy increasingly emphasizes open models, efficiency, sovereignty and specialized performance. Its systems for document intelligence, formal reasoning and coding demonstrate that the most useful model may be the one built for the work rather than the one carrying the highest general intelligence score. For enterprises concerned about control, deployment location or dependence on a small number of American and Chinese providers, Mistral remains strategically relevant even when it does not sit atop the overall leaderboard. The exact order will change, probably before anyone finishes reading this article. The broader pattern will not. OpenAI, Anthropic, xAI, Google, Meta, Microsoft, DeepSeek, Alibaba, Zhipu, MiniMax and Mistral are pursuing different strategies, but nearly all are moving toward models that can reason over longer periods, use tools and take action. The generational comparisons reinforce that point. Sol improves on GPT-5.5, Muse Spark 1.1 advances materially over the original Muse Spark and GLM-5.2 raises the open-weight bar over GLM-5.1. Fable’s four-point gain over Opus 4.8 is smaller than some of the others, but Opus 4.8 was already operating at the frontier. Moving from 56 to 60 at that level is not the same as improving a less capable model by four points. Figure 2. Selected predecessor-to-current gains using current Artificial Analysis v4.1 scores. These comparisons also require caution. Artificial Analysis has revised its index methodology over time, which can cause a model’s current score to differ from the number reported when it launched. Opus 4.8, for example, originally scored 61.4 under an earlier version of the index but scores 56 under the current v4.1 methodology used in this comparison. The only honest way to show generational progress is to compare scores calculated under the same current system. What the chart demonstrates is not that every release produces the same size improvement. It is that competitive pressure is moving the capability curve upward across multiple companies and countries. Better Answers Are Only Part of the Advance During the first few years of generative AI, most improvements were presented as improvements in output. The new model wrote better prose, solved harder math problems, generated more accurate code or hallucinated less frequently. Those advances were real, but the user still had to break the work into pieces and feed those pieces to the model. The user remained the planner, supervisor, integrator and quality-control department. The newest models are beginning to absorb portions of those roles. Instead of requesting one isolated deliverable, I can provide a body of source material and a larger objective. The model can determine what information is present, identify what is missing, conduct additional research, organize the work and create something usable. If it encounters a failure, it has a better chance of diagnosing the problem and trying another approach rather than stopping or repeating the same mistake. The model is not replacing human judgment. I remain responsible for the thesis, the voice, the standards and the decision about whether the result is good enough to publish. I also remain responsible for verifying material facts. These systems can still be wrong, and increased fluency can make an error more persuasive rather than less dangerous. What has changed is the level of work I can entrust to them. A benchmark can tell me that a model solved a certain percentage of coding tasks. It cannot easily measure how much better it is to work with a model that remembers the objective, recognizes when it has drifted and does not require the assignment to be restated every few exchanges. It cannot fully capture the moment when the model stops feeling like a slot machine for good answers and starts feeling like a capable, if imperfect, participant in a project. That is what I see in Sol and Fable. It is also why the advance deserves more than applause. A model that writes a more polished paragraph is useful. A model that can create a plan, operate tools, delegate subtasks, recover from failure and persist until it reaches an objective is more consequential. The same capabilities that make it valuable in a legitimate work environment make it potentially formidable when paired with malicious intent, excessive access or a dangerous objective. We do not need science-fiction speculation about a model waking up and deciding to conquer the world. The immediate concern is concrete. Give a capable agent access to code repositories, communication systems, financial accounts, cloud infrastructure or security tools and the consequences of a poorly defined—or maliciously defined—objective become much larger. The risk does not come from intelligence alone. It comes from intelligence connected to permissions, tools and the ability to act. That should lead organizations to become much more serious about identity, access controls, observability, approval boundaries and the ability to interrupt an agent before one mistake becomes a sequence of them. As models become better at persisting toward an objective, organizations must become better at controlling which objectives they are permitted to pursue. Cybersecurity has taught us this lesson repeatedly: We celebrate what a new technology enables before reckoning with what it enables in the hands of an attacker. Agentic AI will not be an exception merely because the demonstrations are impressive. Sometimes “Good Enough” is the Right Answer The growing power of frontier models does not mean every task should be handed to the largest and most expensive model available. The emerging “good enough” strategy acknowledges that intelligence should be matched to the job. Using maximum reasoning effort for every email, summary or routine classification task is wasteful. Tokenmaxing may produce a marginally better result, but it can also turn AI economics into an obstacle to adoption. Most organizations will use a mix of models. Smaller and faster systems will handle routine work. Specialized models will take on narrow domains. Premium frontier models will deal with ambiguity, orchestrate other systems and manage assignments where judgment matters most. In some architectures, a highly capable model will plan the work while less expensive models execute individual steps. The economics make that approach difficult to ignore. Figure 3. Intelligence Index score versus blended API price per million tokens. This is not the complete cost of finishing a task. The chart does not measure the complete cost of finishing a task. A model that consumes more reasoning tokens, takes more steps or calls more tools can cost considerably more even when its advertised token price appears competitive. It does show the range of strategies emerging in the market. Fable and Sol occupy the maximum-capability end of the field. Meta, Google and several Chinese providers are competing much more aggressively on the relationship between capability and price. That is how frontier intelligence becomes economically usable. There is also a relationship between the two ends of the market. Today’s expensive frontier capability has a habit of becoming tomorrow’s affordable baseline. The leading models establish what can be done. Competition, optimization and smaller architectures determine how cheaply it can be delivered. The ceiling keeps rising while the floor catches up. I have not spent enough time with Grok 4.5, Gemini 3.5, Muse Spark 1.1, the latest Chinese models, Mistral’s current offerings or Microsoft’s agentic systems to judge them from firsthand experience. I will. I do not expect all of them to be equally strong or equally useful. But I expect the bar to keep rising because none of these companies can afford to let a competitor establish a lasting lead in intelligence, agency, speed or cost. For now, I can say that Sol and Fable have moved the needle for me—not in a benchmark or a carefully staged demonstration, but in the work I do every day. They produce better work, understand larger assignments and require less hand-holding. More importantly, they can increasingly help carry a project rather than simply respond to it. The industry is right to pursue efficiency. You do not need a Ferrari to pick up a quart of milk. A Smart car may be cheaper, easier to park and perfectly suited to the errand. But there are roads, races and moments when nothing else will do. The newest frontier models are beginning to show us what happens when the Ferrari can also plan the route, drive itself and decide how to reach the destination. That is exhilarating when the destination is ours. It becomes considerably more serious when it is not.