Dean McCoubrey Co-Founder and Chief AI Strategy Officer of Humaine
TL;DR: GPT-6 Astra and Claude Fable 5.1 launched 48 hours apart in September 2026 at identical list prices. Astra leads on computer use, automation and cost per completed task. Fable 5.1 leads on deep reasoning, long-context knowledge work and cache economics for agentic workflows. Independent benchmarks from Artificial Analysis score both at 53 on the composite intelligence index. There is no overall winner.
That framing matters, because the question most organisations are asking is the wrong one.
The arrival of each new model tends to trigger another migration. People move their work, rebuild their prompts and announce that the previous model has suddenly become obsolete. Most teams still have considerable untapped capacity inside their existing AI subscriptions, and could achieve larger gains by improving their instructions, context, workflow design and quality controls.
The industry has developed a mild case of model-related attention deficit. I call it the magpie effect. Something newer and shinier appears, so we hop towards it before understanding what we already have.
What is the AI magpie effect?
The AI magpie effect is the tendency to chase every new model, tool or feature without first establishing whether it will improve the work. I first described the concept in Daily Maverick in September 2025, warning that C-suite leaders were mistaking spectacle for strategy.
AI platforms encourage this behaviour. New model announcements arrive with benchmark charts, grand claims and carefully selected demonstrations. Within hours, social media is full of declarations about a new winner.
OpenAI launched GPT-6 Astra on 3 September 2026. Anthropic released Claude Fable 5.1 two days earlier, on 1 September. Within a week, organisations were already asking which model they should standardise across their teams.
That question arrives too early for most organisations.
A business should first understand its work, its current capabilities and the limitations it is experiencing. The model should solve an identified problem. Otherwise, the organisation is simply changing engines while the car remains parked.
Most teams still have considerable untapped capacity inside their existing AI subscriptions. They could achieve larger gains by improving their instructions, context, workflow design and quality controls.
Do most people need GPT-6 Astra or Claude Fable 5.1?
Most users probably do not need these models for most of their daily work.
A capable mainstream model, or a model using a medium reasoning setting where that option is available, can manage everyday research, writing, summarisation, meeting preparation, brainstorming and analysis.
I still work regularly in Claude Opus 5 because it performs extremely well across much of my own strategy, research and writing work. Familiarity also matters. A model you understand and direct well can be more useful than a more powerful model you use badly.
Astra and Fable become valuable when the difficulty, length or autonomy of the assignment genuinely requires them. These models are designed for work such as:
- Analysing extensive collections of documents
- Solving complicated coding and technical problems
- Conducting sustained research across multiple sources
- Operating browsers and software autonomously
- Completing multi-stage professional workflows
- Checking and correcting their own work
- Working for extended periods with limited supervision
- Producing complex documents, spreadsheets and digital artefacts
Using a frontier model to rewrite a short email is similar to hiring a senior architect to straighten a picture frame. The architect can certainly do it. The additional capability contributes very little to the result.
What extra capability do Astra and Fable provide?
Astra and Fable provide greater reasoning depth, larger working contexts and a stronger ability to complete long sequences of work.
OpenAI describes Astra as its most capable model for difficult end-to-end professional assignments. It supports computer use, web search, file search, code execution, connected tools and document creation. It also has a context window of approximately one million tokens.
Anthropic positions Fable 5.1 as its most capable generally available model for ambitious, long-running work. It can plan tasks, use tools, recover when a step fails and continue working across multiple applications. Anthropic says it is particularly strong in coding, knowledge work, document analysis and scientific research.
What is a context window? The phrase describes how much information a model can consider during a task. A million tokens represents an enormous quantity of text, potentially equivalent to around 1,500 conventionally formatted pages.
The ability to process that amount of information is useful. It can also create waste when the information has been selected carelessly. A large context window is a capacity, not an instruction to fill it.
Is Claude Fable 5.1 the thinker and GPT-6 Astra the operator?
Claude Fable 5.1 can be understood as the philosopher in the room. GPT-6 Astra behaves more like the operator.
This is an emphasis rather than a hard division. Both models can reason, research, code and use tools.
Fable: depth and sustained reasoning
Fable’s character becomes apparent during deep, sustained assignments. It can examine a large evidence base, maintain a complex argument and investigate the underlying cause of a problem. Anthropic reports that it can run for hours, recover from failed steps and keep users informed as it works. In independent testing by Artificial Analysis, Fable leads on Humanity’s Last Exam (65.0% vs 57.2%) and performs particularly well on long-context reasoning and scientific coding.
Astra: execution and production
Astra has strong reasoning capabilities, with a particular strength in moving from thought into action. It can navigate websites, manipulate software, work across files, test digital experiences and create finished professional outputs. Astra leads on computer use benchmarks (72.6% on OSWorld 2.0), professional document automation and cybersecurity tasks.
For an agency, the division is practical
Fable could analyse interviews, research reports, performance data and client history to identify the strategic problem. Astra could turn the approved answer into a presentation, implementation plan, spreadsheet or working prototype.
Fable asks whether we are solving the right problem. Astra starts opening the applications required to solve it.
What do independent benchmarks say about Astra and Fable?
Independent testing indicates that Astra and Fable are closely matched overall, with meaningful differences across particular kinds of work. According to Artificial Analysis Intelligence Index v4.3, both models score 53 on the composite intelligence index, making this a genuine split decision rather than a clear winner.
| Benchmark | GPT-6 Astra | Claude Fable 5.1 | Edge |
|---|---|---|---|
| Intelligence Index (AA v4.3) | 53 | 53 | Tied |
| FrontierMath Tier 4 | 97.6% | 87.8% | Astra |
| Humanity’s Last Exam (with tools) | 57.2% | 65.0% | Fable |
| GPQA Diamond | 96.0% | 93.7% | Astra |
| AutomationBench | 41.6% | 32.1% | Astra |
| Cost per index task | ~$3.26 | ~$7.63 | Astra |
The pattern: Astra wins on math, science, computer use and cost per task. Fable wins on the hardest reasoning rows and long-context knowledge work.
Treat benchmarks as evidence, not a buying decision. A company should test both models against its own work before standardising on either.
A practical agency evaluation might include a strategic research task, a website audit, a campaign presentation, a spreadsheet analysis and a multi-step account-management workflow. The model should then be assessed on factual accuracy, strategic quality, output quality, time taken, cost, human correction required, compliance with instructions and safe use of permissions.
How much do GPT-6 Astra and Claude Fable 5.1 cost?
The headline API prices are identical. The real cost difference emerges in how each model handles caching and large contexts.
| Cost item | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|
| Input per million tokens | $10 | $10 |
| Output per million tokens | $50 | $50 |
| Cached input per million tokens | $1.00 | $0.25 |
| Requests above 272K input tokens | 2x input, 1.5x output | No surcharge |
Anthropic estimates that Fable 5.1’s lower cache price makes it approximately 25 per cent cheaper than Fable 5 for typical workloads, and up to 45 per cent cheaper for highly agentic work where agent loops constantly re-read large cached contexts. Astra becomes more expensive once a request crosses 272,000 input tokens, at which point the full request is billed at the higher rate.
For subscription users, the cost picture is less visible. Usage allowances can be consumed faster with more capable models or heavier reasoning settings, without a direct per-token bill appearing.
Organisations should monitor cost per completed task, not just token consumption. This includes model usage, employee time, corrections, failed attempts and quality assurance. That is the number that reveals whether a frontier model is earning its place.
Why do autonomous AI models need stronger instructions?
Greater autonomy increases the importance of precise instructions, constraints and stopping points.
An autonomous AI model can decide which steps to take, which tools to use and how long to continue working. This can save an enormous amount of time. It can also produce unnecessary work, consume more tokens or make changes the user did not intend.
“Research this market and prepare a plan” leaves a considerable amount open to interpretation. Every substantial instruction for an autonomous model should define:
- The outcome the model must achieve
- The evidence it may use
- The systems it may access
- The actions it may take
- The actions requiring human approval
- The quality standard for the finished work
- The budget or time available
- The conditions under which it must stop
A stronger instruction would identify the market, approved sources, time period, commercial question, required format, maximum length and decisions reserved for a human. The model also needs to understand what a great outcome looks like. Examples, evaluation criteria and reference materials help it distinguish between technically complete work and genuinely useful work.
Power without direction creates activity. Good operating instructions convert that activity into value.
How can poor context selection increase AI costs?
AI costs rise when we provide excessive context, request unnecessarily long outputs or allow an agent to continue without clear boundaries.
Imagine a project folder containing 100 files. Only six files contain information relevant to the assignment. Uploading the whole folder may cause the model to process a large quantity of irrelevant material. The model has no natural incentive to protect the token budget. It will inspect what it has been given because that appears to be part of the task.
Before beginning expensive work, identify the files that are genuinely relevant, the sections within those files that matter, the period the research should cover, the questions the evidence must answer, the output required and the point at which additional research has declining value. Good context selection improves focus, can reduce cost, and lowers the risk of the model finding a distracting detail and building an elegant answer to the wrong question.
What should businesses be cautious about?
Businesses should pay close attention to permissions, privacy, cost and the reliability of autonomous actions.
OpenAI says Astra has reached its “Critical” cybersecurity capability threshold. The company uses monitoring systems that can pause or stop activity when potentially unauthorised behaviour is detected. Anthropic applies enhanced safeguards to Fable 5.1: certain cybersecurity and biology requests may be routed to another model, and Fable requires 30-day data retention by default, although eligible enterprise customers may qualify for zero data retention or Enterprise Frontier Safeguards.
Any organisation using these models should apply the following controls:
- Apply the minimum necessary permissions
- Keep client environments separated
- Confirm current data-retention terms before deployment
- Test agents away from live systems
- Require approval for consequential actions
- Set cost and time limits on every agentic task
- Keep records of important decisions
- Review outputs for accuracy, confidentiality and bias
Telling an autonomous model what it can do is only part of the job. It must also be told what it cannot do, when it needs permission and when its work is finished.
What is the real lesson from GPT-6 Astra and Claude Fable 5.1?
The real lesson concerns model literacy.
Organisations need to understand the capability they already have, the additional power they are buying and the conditions under which that power creates value. Astra and Fable can complete extraordinary work. They can also be overused, poorly directed and unnecessarily expensive.
The aim is to choose the appropriate level of intelligence, provide precise context, define the destination, control the permissions and understand the cost.
Model selection is becoming an important management discipline. The Financial Times reported that OpenAI claims to have overtaken Anthropic in key capability areas, while independent evaluators show the two as essentially tied. Both claims can be true simultaneously, because they measure different things. The most capable organisations will understand when a frontier model is justified, when a mainstream model is sufficient and when a human should simply do the work.
The industry will continue producing new and shinier objects. We should admire them, test them and resist carrying every one of them back to the nest.
Frequently asked questions
Can a medium-level AI model handle most everyday business tasks?
Yes. A capable mainstream model can manage many common writing, research, summarisation, analysis and administrative tasks. Frontier models become valuable when work requires greater depth, extensive context, sophisticated tool use or long-running autonomy.
How can a business prevent an AI agent from doing too much?
Define its permitted tools, actions, budget, time limit, approval points and stopping conditions. Test the workflow in a controlled environment before allowing access to live client or business systems.
Is the newest AI model always the best choice?
No. The best choice depends on the task, required quality, risk, speed and cost. A familiar and appropriately capable model may produce better commercial results than a more powerful model that is poorly instructed.
Which model is cheaper to run at scale?
It depends on the workload. At the same headline price, Fable 5.1 is significantly cheaper for cache-heavy agentic work ($0.25 vs $1.00 per million cached tokens). Astra is cheaper per completed index task ($3.26 vs $7.63) because it uses fewer output tokens to reach the same score. For requests above 272,000 tokens, Astra applies a surcharge while Fable does not.
What should a business do before adopting AI?
Understand the work you are already doing and where it is breaking down. Identify the specific problem a model would solve. Most teams have significant untapped capacity in their existing tools. The model should follow the diagnosis, not replace it.
How do I use AI without losing our brand quality?
Define what great output looks like before the model starts. Provide examples, evaluation criteria and reference materials. Require human review of all client-facing work. Autonomous models are powerful, but brand judgement, tone and relationships remain human responsibilities.
What is AI theatre and how do you avoid it?
AI theatre is the appearance of AI adoption without the substance. It happens when teams switch models, rebuild prompts and announce new tools without improving the work. Avoid it by measuring outcomes per task rather than model capability, and by mastering what you have before chasing what is new.