Astra vs. Fable: What Are OpenAI and Anthropic Competing For?

We use AI extensively in our own development work. It can produce several technically sound implementations of the same feature, and that is precisely where the human decision becomes clearer. Which version suits the customer? Which interaction asks for too much attention? Which design feels trustworthy? Which compromise will still make sense when the team has to maintain it next year? The model can offer options and evidence. It cannot turn a preference into an objective law.

This matters because frontier releases are usually presented as a ladder: the new model scores higher, so every user has moved one rung closer to a better result. Real work is not that tidy. A stronger model may solve a difficult repository-wide migration and still add flourishes to a simple page that the customer never wanted. It may reason deeply about an instruction while missing the plain request to leave one part untouched. Capability and fit overlap, but they are not identical.

Anthropic's own prompting guidance now documents behavioural differences for Fable 5, including instruction following, effort, long runs, and scaffolding. That does not prove that Claude models generally cannot follow instructions. It does validate the practical point behind user frustration: a model upgrade can require teams to retune instructions that previously worked. When a capable system still needs close direction, the useful question is not only “How intelligent is it?” but “How predictably can we use it?”

Codex users have felt this tension with Astra. OpenAI's current usage guidance does not give a fixed Astra-to-Sol task multiplier: it says allowance use varies with the task, model, and settings. For Plus and Standard Business, its estimates range from 5–45 local messages in a five-hour window for Astra and 15–160 for GPT-6.1 Sol; OpenAI cautions that these are not guaranteed counts and that cloud tasks may use more. That makes the practical access trade-off real, but highly dependent on the work and plan.

Fable has a similar affordability problem at the frontier. As checked on 3 October 2026, Anthropic lists Claude Fable 5.1 at $10 per million input tokens and $50 per million output tokens, and Claude Sonnet 5 at $2 and $10 respectively. These are standard API rates; caching, batch processing, and long-context use have separate terms. Fable may be the correct model for a long, difficult project. It is difficult to justify as the default for work a cheaper model can complete reliably.

This creates an odd market. People who understand the tools and can justify larger subscriptions often squeeze a great deal of value from them. Everyone else learns on the free tier or a smaller plan, exactly where frontier access is limited or metered most tightly. The people who would need experimentation to discover a valuable workflow have the least room to experiment. A model can lead the field and still remain peripheral to most people's working lives.

A monthly subscription makes AI feel like ordinary software. Pay a fixed amount, open the product, and use it. Underneath that interface, every request consumes inference: accelerators run the model, memory holds its weights and working context, networks move data, and data centres supply power and cooling. Long agent sessions add tool calls, repeated context, cached material, and many rounds of generated tokens. The marginal cost is not zero simply because the price is bundled.

As checked on 3 October 2026, Astra and Fable 5.1 list the same headline Standard API rates: $10 per million input tokens and $50 per million output tokens. This is a published rate comparison, not a full bill estimate: OpenAI applies different rates to long-context prompts, caching, Batch, Flex, and Fast, while Anthropic also lists separate cache, batch, and long-context terms. These are premium systems aimed at work where their additional capability can repay the additional cost. OpenAI says Astra can sometimes complete tasks with fewer output tokens, and Anthropic says cheaper cache reads in Fable 5.1 can reduce the cost of typical and highly agentic workloads. Those improvements matter because price per token alone does not tell us price per solved problem.

Still, provider usage guidance and API prices expose more of the underlying economics than a polished chat box does. Providers can subsidise early use and route simple requests to cheaper models, but sustained frontier reasoning has to be paid for somewhere. If the best systems remain expensive to serve, they will stay concentrated among businesses and specialists who can prove a return. That is a distribution limit as much as a pricing decision.

We would like to see the competition widen. Keep improving reasoning, reliability, and safety, but treat affordable inference as a first-class model achievement. A model that is slightly less impressive on an abstract benchmark but available throughout a real workflow may create more value than the leader a user is saving for exceptional prompts.

Lower costs also change what businesses can build. An assistant can review more documents, an agent can check its work, a support workflow can serve smaller customers, and a solo founder can test an idea without treating every retry as a budget event. This is not an argument for unlimited free frontier compute. Somebody pays for the hardware and electricity. It is an argument that reducing the cost of useful intelligence expands the market more directly than adding an ability most customers rarely need.

Both companies publish model and price tiers, but the names, availability, and relative rates change. Compare the current model catalogue and rate card for the exact API or subscription product rather than treating any one lineup as permanent. The labs know that one model cannot optimise intelligence, speed, and price for every job. The next step is to make the capable default tier better enough and cheap enough that model selection stops dominating the user's attention.

Models can become more efficient through better architectures, routing, caching, context management, quantisation, and serving software. But those gains eventually meet physical systems. Memory bandwidth, accelerator utilisation, networking, power, cooling, and the number of chips needed to serve concurrent users all shape the cost of an answer.

That is why OpenAI is building Jalapeño, a purpose-built inference chip with Broadcom. OpenAI describes the project as full-stack work across the model, serving software, chip, memory, and network, and reports early gains in work per watt and latency. We should treat the company numbers as vendor-reported results until broader deployment provides more evidence, but the direction is right: cheaper inference requires hardware and software to be designed together.

The commercial consequence is straightforward. If a provider can produce more reliable completed tasks from the same rack, power budget, and hour, it has more room to lower prices or increase included usage. Hardware efficiency is not a distant infrastructure story. It eventually becomes the difference between an agent a small business can leave running and one it can afford to open only occasionally.