Anthropic argues that internal usage data showed 71% of API customers used exactly two models (almost always Opus + Haiku), with Sonnet appearing mainly in eval scripts and abandoned experiments. The middle tier was a phantom, so collapsing to a two-tier Fable/Mythos scheme with a shared version number better matches how developers actually route traffic.
The editorial frames the rename from poetry forms to narrative forms as an admission that the capability-vs-cost slider with a middle option was the wrong mental model. It endorses Anthropic's read that real traffic patterns bifurcate into cheap-and-fast vs. expensive-and-smart, with nothing meaningful in between.
The editorial highlights that Haiku 4.7 at $0.80/$4 has no direct replacement — Fable 5 with latency_priority: 'low' costs roughly 3× as much for similar latency. Anthropic's suggested workaround (route cost-sensitive workloads through the Batch API for a 50% discount) only approximates Haiku's old pricing and assumes workloads tolerate batch latency, which many production routes do not.
By giving both Fable and Mythos the same version number (5), Anthropic signals that the two tiers are capability-matched siblings rather than independently versioned products. This is meant to eliminate the confusion of cross-tier version mismatches (e.g. Opus 4.7 vs Haiku 4.5) and make model upgrades a single coordinated event.
The editorial notes that deprecated IDs (including date-less aliases) return HTTP 410 Gone on September 9, 2026 — 90 days after announcement. For teams running Opus/Sonnet/Haiku in production with locked-down model IDs, that timeline forces migration, re-evaluation, and re-pricing on a tight schedule with no graceful fallback.
Anthropic published the Fable 5 / Mythos 5 announcement at 9:00 AM PT today, and within four hours the HN thread had crossed 1,200 points. The headline change is structural, not incremental: the three-tier Opus/Sonnet/Haiku lineup is dead, replaced by a two-tier Fable/Mythos scheme that shares a single version number (5) across both models.
The deprecated IDs — `claude-opus-4-7-20260301`, `claude-sonnet-4-7-20260301`, `claude-haiku-4-7-20260301`, plus their date-less aliases — keep serving requests for 90 days. On September 9, 2026 they return HTTP 410 Gone. Anthropic's migration guide pins the recommended swaps as Opus→Mythos 5, Sonnet→Fable 5, and Haiku→Fable 5 with a `latency_priority: "low"` parameter, which is a new field on the messages endpoint.
The new model card pegs Mythos 5 at $18 input / $90 output per million tokens (up from Opus 4.7's $15/$75) and Fable 5 at $2.40 input / $12 output per million (down from Sonnet 4.7's $3/$15). Haiku 4.7 was $0.80/$4 — there is no direct replacement at that price point. Anthropic's blog post argues the new Fable 5, run with `latency_priority: "low"`, hits Haiku-class latency at ~3× the cost, and that customers running Haiku for cost rather than latency should batch through the Batch API for a 50% discount that lands them roughly where Haiku 4.7 was.
The naming shift from poetry forms (Opus, Sonnet, Haiku) to narrative forms (Fable, Mythos) is not just branding. Anthropic is admitting that the three-tier mental model — capability-vs-cost slider with a middle option — was the wrong abstraction for how developers actually route traffic. Internal data Anthropic shared in the post: 71% of API customers used exactly two models in production, almost always Opus + Haiku, with Sonnet showing up only in eval scripts and abandoned experiments. The middle tier was a phantom.
That lines up with what most teams have been seeing. If you're routing on cost, you pick the cheapest model that clears your quality bar — Haiku — and you escalate to Opus only on the requests that fail a confidence check. The middle was a hedge for indecision, not a destination. Anthropic's two-tier reset codifies the actual usage pattern.
The feature matrix is where it gets interesting. Mythos 5 ships a 1M-token context window (up from Opus 4.7's 200K), native multi-step tool use with up to 50 chained calls per request (up from 15), and a new `thinking_budget` parameter that caps reasoning token spend per request — which directly addresses the cost-blowup complaints from Opus 4.7's extended-thinking mode. Fable 5 ships the same 1M context but caps tool chains at 10 and disables extended thinking entirely. Both models share the same tokenizer and the same vision encoder; the difference is purely in the post-training and the reasoning unlock.
The benchmark deltas are smaller than the price deltas suggest: Mythos 5 beats Opus 4.7 by 4.2 points on SWE-bench Verified (72.1 → 76.3) and 6 points on AIME 2025 (84 → 90), but trails GPT-5.1 Ultra by 2 points on both. Fable 5 vs Sonnet 4.7 is a wash on most benchmarks (within 1 point either way) — the value is the 20% price cut and the 1M context, not capability gains. If you were waiting for a generational leap, this isn't one. It's a consolidation release that fixes the pricing curve and the routing model.
Community reaction split predictably. The top HN comment from a Replit infra engineer: "Finally. We've been routing Sonnet → /dev/null for six months." The dissent, from a Cursor backend dev: "Killing Haiku at $0.80 input is going to hurt anyone running cheap autocomplete. Fable 5 + Batch is not the same product — latency math is different." Both are right. The bottom of the price curve genuinely got more expensive for synchronous workloads.
If you have `claude-3-5-sonnet-` or `claude-3-7-sonnet-` or any 4.x model ID hard-coded as a string anywhere — config, env var, prompt template, OpenAI-compatible proxy — grep for it now. The deprecation is a hard cutover on Sept 9; there is no soft-fail to a replacement model, just 410s and a `migration_hint` header pointing at the docs. Three months feels generous until you remember most teams discover their hard-coded model IDs the morning a customer files a P0.
Routing logic needs a rewrite, not a search-replace. The old pattern — cheap model first, escalate on low confidence to mid, escalate again to top — collapses to two tiers, which means your confidence threshold has to move. If you were escalating Haiku→Sonnet at confidence < 0.7 and Sonnet→Opus at < 0.5, the naive port (Fable→Mythos at < 0.7) will route 3-4× more traffic to the expensive model than your old pipeline did, because there's no longer a cheap second-tier catch. Re-tune your thresholds on your eval set before you ship.
The `thinking_budget` parameter on Mythos 5 is the single most important new knob for cost control. Default is unlimited, which means a complex reasoning prompt can silently spend 30K thinking tokens at $90/M before producing output. Set a budget. The Anthropic post recommends 8K as a starting point for code-gen workloads and 2K for classification — both numbers come from their own internal eval set, so adjust upward for your domain. Teams running Opus 4.7 extended thinking should expect their bills to drop 30-50% just from this parameter, before any tier consolidation savings.
The Fable/Mythos naming reset is the first time Anthropic has broken naming continuity since the Claude 2 → Claude 3 jump in early 2024, and it's the first time they've collapsed tiers. Read it as a signal that the "capability tiers as poetry forms" abstraction was holding the company back from making harder pricing and routing decisions, and that the next 18 months of model releases will look more like product-line management than research-lab nomenclature. Expect OpenAI and Google to study the two-tier reception carefully — both are still shipping three-to-five-tier menus that nobody routes against cleanly. If Fable 5 + Mythos 5 lands well, the three-tier model is over industry-wide by mid-2027.
Impressions from testing Fable 5 prior to launch:• My most noticeable immediate jump was in how its frontend design was much more intentionally crafted, and delightful without feeling like 'AI vibe coded'; with better end-user usability too.• In some internal agentic harnesses, it achieved
> In light of the ability of recent models to accelerate their own development, we’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML acceler
From today through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, we’ll remove Fable 5 from those plans. Using it after that will require usage credits. If capacity allows, we’ll extend the included window. After this point—when sufficie
I can’t help but think that there are so many astroturfed comments in here.Seems like a concerted and distributed effort from the entire Anthropic team every time to get this on top of HN.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I've spent enough time with this now in Claude Code (and Claude.ai and Claude Code for web) to have an opinion on Fable 5: it's a beast. I'm throwing some VERY difficult problems at at - things I've been dragging my heels on for months - and it's crunching through them very