Anthropic's Next Models Appear in API as Fable 5 Stalls at 11% of Enterprise Spend

Two unannounced Claude model identifiers — claude-marshmallow-eap and claude-melon-eap — surfaced in Anthropic's developer API and community Discord on August 24, 2026, signaling a new release cycle before the company has acknowledged any upcoming model. The discovery lands the same week that enterprise spending data from payments firm Ramp exposed a harder problem: Anthropic's existing flagship, Claude Fable 5, has been commercially outrun by a cheaper model the company released just a month ago. Together, the two developments sketch a company that is moving fast on new checkpoints while still working out whether its pricing architecture can support a reported $2 trillion IPO valuation.

What the Food Names Signal: Anthropic's EAP Naming Convention

The eap suffix stands for Early Access Program, the label Anthropic attaches to model checkpoints it circulates to a small number of developers before any public release. The food-name convention has precedent: before Claude Fable 5 launched on June 9, developers had spotted claude-fruitcake-eap in the same channel. An earlier checkpoint, claude-honeycomb-eap, appeared during testing that preceded this summer's Sonnet 5 and Opus 5 releases. Anthropic's commercial models use a literary hierarchy — Haiku, Sonnet, Opus, Fable, Mythos — while its internal checkpoints appear to use disposable food codenames that do not reliably predict the commercial name a model eventually ships under.

Early-tester feedback, relayed via developer communities including the X post from Max For AI that first reported the model identifiers, places Marshmallow ahead of Melon in conversational quality. Neither model, according to those assessments, reaches Fable 5's capability tier. Community speculation maps Marshmallow to an Opus-class refresh and Melon to a Sonnet-class iteration, though Anthropic has not confirmed either inference. The presence of two EAP identifiers simultaneously suggests the next release may involve more than one product-line update.

Fable 5's Enterprise Adoption Problem, Quantified

Ramp, which tracks AI spending across roughly 70,000 American businesses, published data showing that Claude Fable 5 has accounted for approximately 11% of what its tracked companies spend on Anthropic tools since the model's June launch — and that figure has remained flat for two consecutive months. According to the Ramp August 2026 AI Index, at the token level, Fable 5 represents only 6% of Anthropic usage by volume, meaning the gap between its share of revenue and its share of actual workload reflects its price premium rather than broad deployment.

The comparison to OpenAI's flagship is pointed. GPT-5.6 Sol, released on July 9, captured 23% of OpenAI's dollar spend and 25% of token volume in the same period. Despite Fable 5 costing approximately $10 per million input tokens — roughly twice Sol's $5 input rate — it generated only about 75% as much July revenue as OpenAI's flagship. Ramp confirmed to the Financial Times that Claude Opus 5 has already overtaken Fable 5 in enterprise spending, less than five weeks after that model launched.

Read more: Anthropic hires Google TPU founder to lead custom chip push

The Opus 5 Cannibalization: A Half-Point Gap at Twice the Price

Claude Opus 5, released July 24 at $5/$25 per million tokens (input/output), is the direct cause of Fable 5's stagnation. On CursorBench 3.2, a coding-focused evaluation, Opus 5 at maximum thinking effort scored 70.0% against Fable 5's 70.5%, according to analysis by Vellum AI. That half-point capability gap comes at a cost-per-task ratio of $8.23 versus $17.32 — more than two-to-one in Opus 5's favor. Dropping one effort level further, Opus 5 at "extra high" actually beats Fable 5 at the equivalent setting (69.3% versus 68.4%) while costing $7.35 against Fable 5's $11.73 per task.

Understanding why this matters requires understanding Anthropic's effort-level architecture. Both Opus 5 and Fable 5 accept a thinking effort parameter — low through max — that controls how much compute the model allocates to extended internal reasoning before responding. For enterprise systems running millions of agent tasks per month, the effort slider is a cost architecture decision, not a minor configuration. Routing those tasks to Opus 5 at medium effort rather than Fable 5 at any setting delivers near-equivalent results at a fraction of the spend.

On Frontier-Bench v0.1, an agentic coding evaluation, Opus 5 leads Fable 5 by nearly ten percentage points — 43.3% against 33.7% — a gap independently confirmed by Vellum AI. Fable 5 retains a narrow lead on SWE-bench Pro and on the dual-use cybersecurity tasks where Anthropic's Mythos-class safety classifiers are specifically designed to manage risk. For most coding and enterprise knowledge work, Opus 5 is the better economic choice. One additional differentiator favors Opus 5: it supports zero data retention agreements, while Fable 5 carries a mandatory 30-day data retention requirement — a meaningful distinction for healthcare, legal, and financial services customers whose compliance frameworks require processing data that never persists on Anthropic's servers.

Fable 5's remaining technical differentiator is its Mythos-class safety architecture. The model runs a real-time classifier on incoming requests in cybersecurity and biology domains; requests that exceed a risk threshold are rerouted to Opus 4.8 for a refusal or restricted response, and Anthropic has reported that classifier triggers occur in fewer than 5% of sessions on average. Anthropic updated those biology-domain classifiers in August 2026, reducing false positives by approximately 85% in testing — a sign that the system's precision continues to improve. For enterprises in regulated industries managing dual-use AI risk, this documented fallback architecture is a genuine differentiator over Opus 5, which carries less restrictive cyber classifiers. The question is whether the set of enterprise buyers who specifically need Mythos-class safety management is large enough to anchor a flagship pricing tier.

The Open-Weight Squeeze: 62% of Tokens at Under 4% of Spend

The internal cannibalization story has a larger industry context. Vercel's AI Gateway — which routes production traffic for more than 200,000 development teams across 200-plus AI providers — shows open-weight models accounting for 62% of token volume as of August 22, according to the platform's live leaderboard, up from 29% in June and roughly 11% in April. Despite handling nearly two-thirds of gateway volume, those models represent less than 4% of spending on the platform. DeepSeek V4 Flash, priced at $0.14 per million input tokens, alone runs 16.9% of all gateway tokens.

Anthropic still commands 65% of Vercel gateway spending on 30% of token volume — a ratio that reflects genuine concentration in high-value, frontier-difficulty workloads. But it also illustrates the ceiling. When a task does not require the best model, development teams send it to the cheapest one that clears their quality threshold. That threshold rises with each open-weight release cycle, compressing the category of work that justifiably demands a $10-per-million-token model.

What Marshmallow and Melon Might Change — and What They Cannot

If Marshmallow is an Opus-class iteration, it arrives into a market where the current Opus already outperforms Fable 5 on most benchmarks at half the price. A stronger Opus improves Anthropic's competitive position, particularly against open-weight challengers in the mid-tier, but it deepens the cannibalization dynamic at the premium tier rather than resolving it. If Melon maps to a Sonnet-class refresh, it enters a tier where Sonnet 5, permanently priced at $2/$10 per million tokens following Anthropic's August 10 pricing decision, already handles the majority of everyday developer tasks cost-effectively.

The stakes extend to Anthropic's capital structure. Multiple Anthropic investors are reportedly modeling a public listing valued at $2 trillion or more, more than double the company's $965 billion post-money valuation from its Series H funding round, according to the Financial Times. The narrative behind that figure requires Anthropic to demonstrate not just that it has the most technically sophisticated models, but that it can translate that capability into durable pricing power in enterprise markets. Fable 5's first two months — high benchmark performance, stagnant revenue share — complicate that story. The comparison that matters most to investors is not Fable 5 against Opus 5 or against GPT-5.6 Sol; it is Fable 5's 11% spending share against the broader expectation that a premium-tier product generates premium-tier revenue concentration.

The structural challenge the Ramp and Vercel data describe is not one that a new model tier automatically solves. Each time Anthropic ships a model that nearly matches its prior flagship at half the cost, it signals — implicitly and numerically — that the prior flagship was priced above what its marginal capability advantage justified. Investors pricing a potential $2 trillion public valuation are weighing two simultaneous signals: 43.5% of US businesses tracked by Ramp paid for Anthropic subscriptions or tokens in July, which reflects real platform breadth; and Fable 5 stagnated at 11.4% of Anthropic spending within two months of launch, which raises questions about premium-tier pricing power. The next milestone to watch is not when Marshmallow and Melon ship — it is whether, when they do, Anthropic arrives with a pricing structure that addresses the cannibalization pattern rather than extending it.

Read more: Anthropic targets record IPO as Claude revenue surpasses $65 billion