Fable 5.1 Slashes Cache Pricing 75%, Muse Spark 1.3 Storms the Mid-Tier: A Flagship Reshuffle in One Week

In the first week of September, three major labs made their moves within 48 hours: on September 2, Anthropic released Claude Fable 5.1 (alongside Mythos 5.1) and cut its prompt-cache read price by 75%; Meta followed with Muse Spark 1.3, holding the previous generation's pricing while eyeing the coding market; and on September 3, OpenAI pushed flagship pricing to $10/$50 with GPT-6 Astra. One cut, one hold, one hike — the flagship landscape was reshuffled in a single week.

Pricing at a Glance

For reference, GPT-5.6 Sol — the model Astra replaces — was priced at $5 input / $30 output, while Muse Spark 1.3 costs one-eighth as much as a flagship.

Anthropic: The Cache War Begins in Earnest

Fable 5.1 keeps exactly the same input and output prices as Fable 5 ($10/$50). The real weapon is cache reads: down from $1 to $0.25 per million tokens. Anthropic's own estimate is that a typical AI application workload sees total costs drop by about 25%, and agent-style tasks with long context and many turns can save up to 45%.

Take that number with a grain of salt: how much you save depends on the share of cache reads in your workload. For traditional one-question-one-answer usage, cache reads are a small fraction and the difference is barely noticeable. But for coding agents and deep research — workloads that repeatedly read the same large context — cache reads often account for more than 70% of total tokens, and a 75% cut in cache-read pricing can literally halve your bill. One outlet measured a long-running task at $2.72, noticeably cheaper than the same task on Astra.

The timing is worth savoring, too: on the very same day OpenAI raised flagship prices with Astra, Anthropic chose to concede on the "hidden cost." The sticker prices are identical at $10/$50, but once you actually run agent workflows, Fable 5.1's total cost comes out significantly lower — pricing competition has moved from the show floor to the bill.

Meta: Taking the Mid-Tier at One-Eighth the Price

Muse Spark 1.3 keeps exactly the same pricing as 1.2 ($1.25 input / $0.15 cache read / $4.25 output) while raising its coding capability to what Meta claims is "beyond GPT-5.6 Sol," with 1M context included. This is Meta's fourth model update in five months — a cadence that is clearly about catching up.

$1.25/$4.25 is a carefully chosen price point: below even GPT-5.6 Terra ($2/$12), with output costing less than a tenth of flagship pricing. For high-volume code review, document processing, and lightweight agent tasks, model quality at this tier is already "good enough," and Meta itself called the pricing "aggressive." Meta also extended its contributor tier — let Meta use your submitted content to improve its models in exchange for a lower price — an option that has proven popular with the developer community.

OpenAI: Raising the Flag and Setting the "AGI Tier" at $10/$50

We covered GPT-6 Astra in detail in our previous post: at $10/$50 it costs 2.5 times as much as GPT-5.6 Sol, ships with 1.05M context, and output rises to $75 once input exceeds 272K tokens. Interestingly, this price exactly matches Claude Fable 5.1 — the two flagships have reached a silent consensus on the "AGI tier," and $10/$50 is becoming the industry's baseline price for top-tier models.

Viewed side by side, the strategic split is stark: OpenAI raises prices to plant its flag and bank on the "strongest model" premium; Anthropic holds the sticker price flat while quietly courting agent developers with cache discounts; Meta bypasses the frontal battle entirely, harvesting cost-sensitive, high-volume workloads at one-eighth the price.

How to Route Across Tiers for the Lowest Cost

In this new landscape, routing by task difficulty is worth more than ever:

• Lightweight preprocessing (classification, extraction, format conversion): Muse Spark 1.3 is the new price-performance king — $1.25/$4.25 with 1M context has virtually no competition; budget-sensitive batch jobs can stay on GPT-5.6 Luna.
• Mid-tier workhorse (everyday coding, RAG Q&A): GPT-5.6 Sol / Terra remain the balanced picks; if your workflow has high cache hit rates, go straight to Fable 5.1 — with cache reads at $0.25, flagship pricing may not cost much more than mid-tier.
• Flagship reasoning (complex inference, long-running agents): Fable 5.1 matches GPT-6 Astra on price, but in agent scenarios its cache advantage makes real bills 30–45% lower — a sensible default flagship; switch to Astra only when you need its exclusive 1.05M context.

Prompt Caching pairs especially well with this: put stable prefixes — system prompts, tool definitions, long documents — at the front of your requests. The higher your cache hit rate, the more of Fable 5.1's price cut you keep.

MuiRouter Is Already Up to Date

MuiRouter has launched Claude Fable 5.1 and Muse Spark 1.3 at official pricing (Fable 5.1 cache reads at $0.25, Muse Spark 1.3 at $1.25/$4.25); GPT-6 Astra was integrated yesterday. All three are available in the Playground, with the model list and billing prices updated in real time.

With MuiRouter's multi-tier routing, you can dispatch each class of request to the right tier — and in a market where flagship prices converge while the mid-tier attacks, it remains the only controllable way to keep costs at rock bottom.