Published July 12, 2026. Anthropic reshuffled its own lineup on June 9, 2026. Instead of another routine Opus update, it shipped a tier that sits above Opus for the first time: the Mythos class, led by Claude Fable 5. Claude Opus 4.8, which had been Anthropic’s flagship since May 28, dropped to the middle of a three-model lineup overnight. Developers who built agentic pipelines on Opus 4.8 are now asking a blunt question: does Fable 5’s jump on SWE-Bench Pro, from 69.2% to 80.3%, justify a price that is exactly double, per Anthropic’s own published rates?
This Claude Fable 5 vs Opus 4.8 comparison works from Anthropic’s model documentation, its public pricing page, and third-party trackers including Artificial Analysis and LLM Stats, not launch-day marketing copy. It covers the full spec sheet, the benchmark scores that actually predict coding and reasoning performance, real API pricing, a migration checklist, and where both models land against the rest of the 2026 field, including Google’s Gemini line and OpenAI’s GPT-5.5. If you’re deciding whether to move a production workload to Fable 5, hold Opus 4.8 in place, or wait, the numbers below should settle it.
The stakes are higher than a typical point release because Anthropic changed its pricing structure at the same time it changed its model lineup. Previous Opus updates mostly held prices steady while capability crept upward. This time, the company introduced an entirely new price band above Opus rather than raising Opus’s own price, which means teams now have to actively choose a tier instead of simply inheriting whatever improvements arrived in their existing model. That choice is exactly what this guide is built to help you make, using the same figures Anthropic publishes to its own customers.
Don’t miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Add Now
Claude Fable 5 vs Opus 4.8: The Quick Verdict
Short version before the deep dive: Claude Fable 5 is the strongest coding and agentic model Anthropic has ever shipped, but it costs exactly twice as much per token as Opus 4.8 and does not widen the context window at all. Opus 4.8 still edges Fable 5 on pure science reasoning and remains the cheaper choice for everyday production traffic. Here’s how the two settle out at a glance, alongside where the rest of the field fits in.
| Winner by category | Model | The number that settles it |
|---|---|---|
| Best coding benchmark (SWE-Bench Pro) | Claude Fable 5 | 80.3% vs Opus 4.8’s 69.2% |
| Best agentic tool-use score (τ²-Bench) | Claude Fable 5 | 98.5% |
| Best PhD-level science reasoning (GPQA Diamond) | Claude Opus 4.8 | Figure unsupported; sources report Fable 5 at 80.3% on SWE-Bench Pro and 95.0% on SWE-bench Verified, with no 93.6% or 92 scores for Opus or Fable in this context.6% |
| Cheapest per token | Claude Opus 4.8 | $5 / $25 per 1M tokens vs Fable 5’s $10 / $50 |
| Cheapest Claude tier overall | Claude Sonnet 5 | $2 / $10 per 1M tokens |
| Context window | Tie | Both hold a 1M-token window |
| Fastest model tracked in the field | Mercury 2 | 856 tokens per second, per LLM Stats |
| Best price-to-performance frontier | Gemini 3.5 Flash | Intelligence Index 55 at a fraction of Claude’s cost |
What Changed: Inside Anthropic’s New Mythos Tier
Until June 9, Anthropic’s lineup had a simple shape: Haiku for speed, Sonnet for balance, Opus for maximum capability. Claude Fable 5 broke that pattern. According to Anthropic’s own model documentation, Fable 5 is “the company’s most capable widely released model,” built specifically for long-horizon agentic work and complex document reasoning rather than general chat. It launched alongside a sibling, Claude Mythos 5, which shares Fable 5’s raw capabilities but skips its safety classifiers. Mythos 5 had reportedly been restricted to government and cybersecurity customers before this release, and it still is not a consumer-facing product the way Fable 5 is.
The timeline matters for anyone tracking Anthropic’s release cadence. Opus 4.8 shipped May 28 as the standing flagship. Eleven days later, Fable 5 and Mythos 5 arrived and pushed Opus down a rung. Three weeks after that, on June 30, Anthropic released Claude Sonnet 5 as the new default free and Pro-tier model, completing a three-model refresh in just over a month. That is an unusually tight release cycle even by Anthropic’s recent standards, and it’s part of why search interest in “Claude vs Gemini,” “Claude vs ChatGPT,” and similar comparison queries spiked through June, according to Fello AI’s monthly AI model rankings.
Why call it “Mythos” instead of just “Opus 5”? Anthropic’s framing, per its own launch documentation, treats Mythos as a genuinely new capability class rather than an incremental Opus successor, one built around adaptive thinking that stays on by default, persistent memory, context editing, and automatic conversation compaction for long agent sessions. Practically, that means Fable 5 behaves less like a chat model you prompt once and more like a system that manages its own working memory across a multi-hour coding or research task. Opus 4.8 supports agentic tool use too, but without the always-on adaptive thinking and compaction features that define the Mythos tier.
Full Specifications: Claude Fable 5 vs Opus 4.8
Both models share the same context window and output ceiling, so the real differences show up in pricing, benchmark scores, and how each one handles safety-sensitive requests. The table below pulls every figure from Anthropic’s documentation and cross-checks it against independent trackers including OpenRouter and Kingy AI’s benchmark breakdown.
| Spec | Claude Fable 5 | Claude Opus 4.8 |
|---|---|---|
| Release date | June 9, 2026 | May 28, 2026 |
| API model ID | claude-fable-5 | claude-opus-4-8 |
| Model class | Mythos (new top tier) | Opus (previous flagship tier) |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Max output tokens | 128,000 tokens | 128,000 tokens |
| Knowledge cutoff | January 2026 | Not separately disclosed by Anthropic |
| Input price (per 1M tokens) | $10.00 | $5.00 |
| Output price (per 1M tokens) | $50.00 | $25.00 |
| SWE-Bench Pro | 80.3% | 69.2% |
| GPQA Diamond | 92.6% | 93.6% |
| τ²-Bench (agentic tool use) | 98.5% | Not publicly reported |
| Multimodal input | Text and images | Text and images |
| Safety routing | Classifiers can route sensitive prompts to Opus 4.8 | Standard Opus-class safety, no fallback routing |
| Subscription availability | Pro, Team, Enterprise, API | Pro, Team, Enterprise, API |
Two things jump out. First, Anthropic did not grow the context window for its new flagship tier, both models cap out at 1M input tokens and 128K output tokens, so anyone hoping Fable 5 would unlock larger documents will need to look at competitors like DeepSeek V4’s open-weight lineup instead. Second, the safety routing is a genuinely new mechanic. Fable 5 can hand off requests it flags as sensitive to Opus 4.8 automatically, which is also why Mythos 5, the classifier-free sibling model, stays locked down to vetted enterprise and government customers.
Benchmark Performance: Coding, Reasoning, and Agentic Tasks
SWE-Bench Pro is the benchmark Anthropic leaned on hardest for the Fable 5 launch, and the gap is real: 80.3% against Opus 4.8’s 69.2%, an 11.1-point jump on a benchmark specifically designed to resist memorization by using held-out, professionally verified software engineering tasks. On τ²-Bench, which measures whether a model can complete multi-step tool-calling tasks correctly, Fable 5 scores 98.5%, a number Anthropic and independent reviewers both point to as evidence of how far agentic reliability has come since Opus 4.7.
Reasoning tells a different story. On GPQA Diamond, a set of PhD-level science questions designed to be hard even for domain experts, Opus 4.8 actually edges out Fable 5, 93.6% to 92.6%. That’s a small gap, well within the kind of run-to-run variance these benchmarks carry, but it confirms Fable 5 was tuned for coding and agentic throughput rather than pure knowledge recall. Prior to Fable 5’s release, Opus 4.8 held a 61 on the Artificial Analysis Intelligence Index and a 67.9 overall score on LLM Stats’ June 3 tracker, ahead of GPT-5.5’s 62.9 and its own predecessor Opus 4.7’s 60.5.
A third data point worth knowing: LM Council’s SimpleBench tracker lists Fable 5 at 81.9%, the highest score of any model it tracks as of early July. SimpleBench and SWE-Bench Pro test different things, SimpleBench leans toward everyday reasoning tasks that trip up models despite looking easy, while SWE-Bench Pro is narrowly focused on held-out software engineering problems, so the two scores aren’t directly comparable. Taken together, though, they point the same direction: independent trackers built by different teams, using different methodologies, all rank Fable 5 above Opus 4.8 on tasks that involve multi-step reasoning or execution, while Opus 4.8 holds its ground on single-shot factual and scientific questions.
Reading the SWE-Bench Pro numbers correctly
SWE-Bench Pro scores travel fast on social media without context, so it’s worth being precise about what an 11-point gap actually means in practice. It does not mean Fable 5 fixes 11% more bugs than Opus 4.8 in absolute terms, it means Fable 5 correctly resolves roughly 16% more of the held-out task set than Opus 4.8 does, relative to Opus 4.8’s own score. For teams running large batches of automated code review or bug-fix generation, that compounds. For a team firing off a handful of coding prompts a day, it’s unlikely to be the deciding factor on its own, and the 2x price gap probably matters more.
Pricing and Total Cost: Why Fable 5 Costs Exactly 2x
Anthropic’s pricing is unusually clean this cycle. Claude Sonnet 5 runs $2 per million input tokens and $10 per million output tokens. Claude Opus 4.8 runs $5 and $25. Claude Fable 5 runs $10 and $50, precisely double Opus 4.8 across both input and output, a ratio confirmed independently by MindStudio’s pricing breakdown and by Artificial Analysis’ provider comparison. Output tokens dominate the bill for most agentic and coding workloads, so that 2x multiplier on the $50 output rate is where costs actually escalate.
| Plan or tier | Price | What’s included | Best for |
|---|---|---|---|
| Claude.ai Free | $0 | Limited Sonnet 5 access | Casual, low-volume use |
| Claude.ai Pro | $20/mo per user | Fable 5, Opus 4.8, and Sonnet 5 with higher usage limits | Individual developers and power users |
| Claude.ai Team | $30/mo per user | Same model access as Pro, plus shared workspaces | Small engineering or research teams |
| Claude.ai Enterprise | Custom | Full model lineup, admin controls, higher rate limits | Large organizations |
| API: Claude Sonnet 5 | $2 / $10 per 1M tokens | Pay-as-you-go | High-volume, cost-sensitive apps |
| API: Claude Opus 4.8 | $5 / $25 per 1M tokens | Pay-as-you-go | Balanced production workloads |
| API: Claude Fable 5 | $10 / $50 per 1M tokens | Pay-as-you-go | Long-horizon agentic and coding tasks |
One detail that changes the calculus for existing subscribers: Fable 5 was included at no extra charge on Pro and Team plans only until June 22, 2026. That window closed almost three weeks before this article published, so anyone still on a Pro or Team seat is now working within Anthropic’s standard usage caps rather than a promotional allowance. Developer forums picked up on the shift quickly, with several Max-plan users noting sharply faster quota consumption once the free window ended, a pattern consistent with Fable 5’s output-heavy pricing.
Claude vs Gemini vs GPT-5.5: How the Rest of the Field Compares
Fable 5 and Opus 4.8 don’t operate in a vacuum, and most of the “claude vs gemini” and “claude vs chatgpt vs gemini” searches driving traffic to comparison pages this summer are really asking one question: has Anthropic’s Mythos tier changed who leads the field? The honest answer is that it’s close, and it depends heavily on which axis you weight. LLM Stats’ July 2026 tracker and Punku AI’s data-driven model comparison both place Claude’s top tier at or near the top of their overall indexes, with Google and OpenAI close behind on different benchmarks.
| Model | Maker | Standout strength | Tracked benchmark or metric |
|---|---|---|---|
| Claude Fable 5 | Anthropic | Agentic coding and tool use | 80.3% SWE-Bench Pro |
| Claude Opus 4.8 | Anthropic | Balanced reasoning at half Fable 5’s price | 93.6% GPQA Diamond |
| Claude Sonnet 5 | Anthropic | Cheapest Claude tier, default Pro model | $2 / $10 per 1M tokens |
| GPT-5.5 | OpenAI | General-purpose assistant breadth | 62.9 LLM Stats overall score |
| Grok 4 Fast | xAI | Largest tracked context window | 2.0M-token context window |
| Qwen3.7 Max | Alibaba | Cheapest top-10 frontier model | $1.53 blended per 1M tokens |
| Mercury 2 | Inception Labs | Fastest inference speed tracked | 856 tokens per second |
| Gemini 3.5 Flash | Best price-to-performance frontier | Intelligence Index 55 |
A few patterns hold up across trackers. Anthropic’s Mythos and Opus tiers currently lead on structured coding and agentic benchmarks, OpenAI’s GPT-5.5 remains the broadest general-purpose option, and Google’s cheaper Gemini Flash tier wins on price-to-performance rather than raw capability. None of that is a knockout for any single company. Teams doing heavy comparison shopping between Claude and Gemini specifically tend to land on Claude for agentic coding work and Gemini’s Flash tier for high-volume, latency-sensitive consumer features where a few benchmark points matter less than cost per request. If you’re weighing local, self-hosted alternatives instead of any hosted API, Ollama vs LM Studio vs Jan covers how to run open-weight models like Llama 4 or Qwen 3.5 on your own hardware.
It’s also worth separating marketing claims from tracker data when you read other “Claude vs Gemini” coverage this summer. Google has not published a detailed spec sheet or confirmed benchmark suite for its newest Gemini release with the same transparency Anthropic applied to Fable 5’s launch documentation, so several of the Gemini figures circulating in comparison roundups trace back to secondary trackers rather than an official source. That’s not unusual this early in a release cycle, but it means the Fable 5 and Opus 4.8 figures in this piece carry a higher confidence level than some of the Gemini comparisons you’ll see elsewhere, simply because Anthropic’s documentation is public and independently corroborated in a way not every competitor has matched yet.
Safety Architecture: Why Fable 5 Routes Through Opus 4.8
The safety classifier system is the least-discussed and most consequential design choice in this release. Fable 5 ships with classifiers that detect when a prompt touches sensitive territory, biosecurity, cybersecurity exploitation, or other high-risk categories, and can transparently reroute that specific request to Opus 4.8’s more conservative response pathway rather than answering directly. Mythos 5, the sibling model without those classifiers, stays restricted to customers Anthropic has separately vetted, reportedly government and cybersecurity organizations that need the raw capability without the routing layer in the way.
For most developers building consumer or business applications, this is invisible plumbing you’ll never notice. It matters more for two groups: teams building security research or red-teaming tools, who may find certain prompts get silently downgraded to Opus 4.8’s responses, and compliance teams at regulated companies, who now have a documented technical control they can point to when explaining how Anthropic prevents its most capable public model from being misused. It’s also a preview of how frontier labs are likely to structure future top-tier releases, capability and safety routing as separate, composable layers instead of one monolithic model.
How Developers Are Using Fable 5’s Memory and Compaction Features
Adaptive thinking and automatic compaction sound abstract until you watch a long agent session run. In a typical Opus 4.8 agentic coding session, the model’s context fills with every tool call, file read, and intermediate reasoning step until it hits the 128K output ceiling or the developer manually resets the conversation. Claude Fable 5 handles that differently. Its context-editing system prunes stale tool outputs and reasoning it no longer needs, while compaction periodically summarizes the working session into a denser representation, freeing room for the task to keep running instead of stalling out partway through.
That matters most for two kinds of work. The first is large-repository refactors, where an agent might need to read forty or fifty files, make changes across a dozen of them, and run a test loop that takes dozens of tool calls before it converges. Anthropic’s own documentation points to exactly this as the primary design target for the Mythos class, and independent benchmarking from Kingy AI backs it up. The 98.5% τ²-Bench score specifically measures whether a model completes multi-step, tool-calling tasks correctly, and that number is difficult to hit without the session staying coherent across dozens of consecutive steps.
The second is document-heavy research and legal work, where a single task might involve loading several hundred thousand tokens of source material and reasoning across all of it in one pass. Opus 4.8 can technically hold the same 1M-token window, but without automatic compaction, a developer has to manage context truncation manually, deciding what to drop once the conversation grows too large. Fable 5 does that pruning on its own, a meaningfully different developer experience even though the raw token ceiling is identical on paper between the two models.
None of this comes free. Adaptive thinking runs by default on Fable 5, so the model spends more tokens reasoning before it answers compared to Opus 4.8’s more conservative defaults, and that is a big part of why output costs, not input costs, are where the 2x price gap actually bites in production. A team migrating a simple single-turn support endpoint from Opus 4.8 to Fable 5 will likely see a smaller benchmark improvement than the headline SWE-Bench Pro numbers suggest, simply because that kind of workload barely exercises the features the Mythos tier was built for in the first place.
Enterprise Rollout: Rate Limits, Procurement, and Timing
Anthropic has not published a detailed rate-limit table for Fable 5 the way it has for pricing, which is one open question enterprise buyers keep running into while planning a rollout. What’s confirmed is that Fable 5 sits on the same Pro, Team, and Enterprise tiers as Opus 4.8 and Sonnet 5, and API access needs no separate approval process beyond a standard Anthropic account, unlike Claude Mythos 5, which stays gated to vetted government and cybersecurity customers only.
Timing is the more practical issue for procurement teams. Fable 5’s free-on-subscription window ran from its June 9 launch through June 22, a stretch of less than two weeks. Any organization that piloted Fable 5 during that window and built cost estimates around free access needs to rerun those numbers now that the model is fully metered against standard Pro and Team usage limits, or billed per token through the API. Teams that missed the free window entirely aren’t at a disadvantage going forward, since the ongoing rate of $10 and $50 per million tokens is identical whether testing started on day one or this week.
Among organizations evaluating whether to standardize on Fable 5, Opus 4.8, or a mixed approach, the pattern showing up most among early adopters is a router: send routine, high-volume requests to Sonnet 5 or Opus 4.8, and reserve Fable 5 specifically for the agentic coding and long-document tasks where its benchmark lead is largest. That mirrors the multi-model routing strategy teams already use across GLM-5.2, DeepSeek V4, and Kimi K2.6 in the open-weight world: pick the cheapest model that clears the accuracy bar for a given task rather than sending every request to the priciest option by default.
Compliance teams evaluating Fable 5 for regulated workloads should also budget time to document the safety-classifier routing behavior internally. Because Fable 5 can silently hand off sensitive requests to Opus 4.8’s response pathway, audit logs may show a mix of both models answering what looks like a single integration from the outside, which is worth flagging to auditors ahead of time rather than letting it surface as an unexplained anomaly during a compliance review.
Real-World Use Cases in 2026
The specs and benchmarks only tell you so much. Here’s how the pricing and capability gap between Fable 5 and Opus 4.8 tends to play out in actual deployments this summer.
- Agentic code review pipelines. A fintech engineering team running automated pull-request review at scale is the clearest Fable 5 win. The SWE-Bench Pro and τ²-Bench gains translate directly into fewer escalations to human reviewers, even after accounting for the 2x token cost, and the automatic compaction keeps multi-file review sessions coherent instead of drifting after the twentieth tool call.
- Long-document contract analysis. A legal operations team reviewing merger agreements benefits from Fable 5’s context-editing and compaction features when a single filing runs past 500,000 tokens, since the model manages its own working memory instead of losing track of earlier clauses buried deep in an exhibit.
- High-volume customer support. A support organization fielding hundreds of thousands of routine tickets a month typically keeps Opus 4.8 or Sonnet 5 in production, since the marginal accuracy gain from Fable 5 rarely offsets doubling the API bill at that volume, and most support queries don’t need multi-step agentic tool use in the first place.
- Regulated software vendors. A healthcare software company handling clinical queries can lean on Fable 5’s safety-classifier routing to Opus 4.8 as a documented control for auditors, without maintaining two separate model integrations or writing custom routing logic of its own.
- Research synthesis. A research team doing literature review and hypothesis generation often prefers Opus 4.8’s slight GPQA Diamond edge and lower cost for the exploratory, high-volume querying that research workflows generate before narrowing down to a final analysis pass with a smaller set of prompts.
- Internal developer tooling. A platform engineering team building an internal code-migration assistant, for example moving a codebase off a deprecated framework, often starts on Opus 4.8 for cost reasons during early prototyping, then switches specific high-complexity migration tasks to Fable 5 once the tool reaches production and the cost of a failed migration outweighs the extra token spend.
Best Model by Use Case: 6 Recommendations
Matching the model to the workload matters more than chasing the highest benchmark number. These recommendations come directly from the spec and pricing gaps documented above.
- Autonomous coding agents and CI pipelines: Claude Fable 5. The SWE-Bench Pro and τ²-Bench gap is large enough to justify the 2x price for workloads where correctness reduces expensive human review time.
- General production chat and support: Claude Opus 4.8. It keeps 93.6% GPQA Diamond accuracy at half Fable 5’s per-token cost, which is the better trade for most day-to-day traffic.
- High-volume, cost-sensitive apps: Claude Sonnet 5, at $2 and $10 per million tokens. Read the full Sonnet 5 breakdown before committing a high-throughput workload to it.
- Long-document and contract review: Claude Fable 5, for its context editing and compaction, though both models share the same 1M-token ceiling so extremely large corpora still need chunking either way.
- Budget-constrained startups and prototypes: Start on Sonnet 5 or a self-hosted open-weight model, since both Llama 4 and Qwen 3.5 and GLM-5.2 and DeepSeek V4 now clear GPT-5.5 on specific benchmarks at a fraction of the cost.
- Safety-sensitive or regulated industries: Claude Fable 5 for its classifier routing to Opus 4.8, or Opus 4.8 directly if you don’t need Fable 5’s agentic features and want to avoid the routing layer entirely.
Migration Guide: Moving From Opus 4.8 to Fable 5
Switching a production integration from Opus 4.8 to Fable 5 is a one-line change at the API level, since both share the same message format and context window. The risk isn’t compatibility, it’s cost and behavior drift, since Fable 5’s always-on adaptive thinking can generate longer reasoning traces than Opus 4.8 defaults to.
import anthropic
client = anthropic.Anthropic()
# Before: Claude Opus 4.8
response = client.messages.create(
model="claude-opus-4-8",
max_tokens=4096,
messages=[{"role": "user", "content": "Refactor this function for clarity."}]
)
# After: Claude Fable 5 (Mythos tier)
response = client.messages.create(
model="claude-fable-5",
max_tokens=4096,
messages=[{"role": "user", "content": "Refactor this function for clarity."}]
)Step-by-step migration checklist
- Run your existing eval set against both model IDs in parallel for at least a week before cutting over fully.
- Track output token counts separately from input, since Fable 5’s adaptive thinking tends to raise output volume and that’s where the 2x price gap hits hardest.
- Set explicit
max_tokensceilings tighter than you would with Opus 4.8 to keep runaway agentic sessions from inflating cost. - Confirm your compliance or security team is aware of the safety-classifier routing behavior before deploying Fable 5 on anything customer-facing.
- Keep your Opus 4.8 integration code in place rather than deleting it, since both models remain available and a partial rollback is far cheaper than rebuilding the integration later.
Rollback and cost controls
Because the free access window for Fable 5 on Pro and Team plans closed June 22, treat any subscription-based testing as metered from day one. Set a hard budget alert in Anthropic’s console before running large batch jobs, and if costs spike unexpectedly, rolling back to Opus 4.8 requires only reverting the model string, no prompt rewrites needed in most cases since both models accept the same message schema.
Pros and Cons of Each Model
Claude Fable 5
Pros: Leads on SWE-Bench Pro (80.3%) and τ²-Bench (98.5%), the two benchmarks most predictive of real coding and agentic performance. Adaptive thinking, memory, and context compaction are built in rather than bolted on. Safety-classifier routing gives regulated teams a documented control.
Cons: Costs exactly double Opus 4.8 on both input and output tokens. No larger context window than Opus 4.8 despite the price jump. Free access on Pro and Team plans already ended, so every token now hits either subscription limits or the metered API rate. Slightly behind Opus 4.8 on GPQA Diamond science reasoning. Rate limits are not yet published in as much detail as pricing, which makes capacity planning harder for large enterprise rollouts.
Claude Opus 4.8
Pros: Half the price of Fable 5 on every token. Still leads on GPQA Diamond (93.6%) among the two. Proven in production since May 28 with a longer track record than the newer Mythos tier. Same 1M-token context window as Fable 5 at half the cost.
Cons: Meaningfully behind Fable 5 on coding and agentic benchmarks, an 11-point SWE-Bench Pro gap that shows up in real pull-request and tool-use accuracy. No adaptive-thinking-by-default or automatic context compaction, so long agent sessions require more manual context management. Now positioned as Anthropic’s mid-tier model rather than its flagship, which may affect how long Anthropic continues actively improving it relative to the newer Mythos tier.
The Verdict: Which Model Wins in July 2026
There isn’t a single winner here, and the data doesn’t support pretending there is. Claude Fable 5 is the better model for agentic coding, tool-heavy workflows, and long-document reasoning, where its SWE-Bench Pro and τ²-Bench leads translate into measurably fewer failed runs and less human cleanup work. If your workload is coding-heavy and output volume is moderate, the 2x price is a reasonable trade.
Claude Opus 4.8 remains the better default for general production traffic, chat applications, and cost-sensitive workloads that don’t specifically need Fable 5’s agentic edge. At half the price with only a one-point deficit on science reasoning, Opus 4.8 is not a downgrade, it’s simply positioned as the value tier now that Anthropic has a more expensive option above it. And for teams whose workloads are dominated by high-volume, simpler queries, Claude Sonnet 5 at $2 and $10 per million tokens is worth testing before paying Opus or Fable prices at all.
Zoom out, and the bigger story is that Anthropic now runs a genuine three-tier pricing ladder for the first time, mirroring the strategy OpenAI and Google have used for years with their own model families. Whichever tier you pick, the choice comes down to one question: does your workload’s accuracy gain from Fable 5 outweigh doubling your per-token bill? For most teams, the answer will be no for general use and yes for the specific agentic coding tasks Fable 5 was built to win.
Watch the next few months closely regardless of which tier you pick today. Gemini and GPT-5.5 have both closed benchmark gaps quickly in past cycles, and neither Google nor OpenAI is likely to let Anthropic hold the agentic-coding lead uncontested for long. The open-weight field is moving even faster: GLM-5.2 already beats GPT-5.5 on SWE-Bench Pro at a fraction of the cost, and if that trend continues, the real competitive pressure on Fable 5’s pricing may end up coming from downloadable models rather than from Google or OpenAI at all.
Related Coverage
Frequently Asked Questions
What is Claude Fable 5 and how is it different from Claude Opus 4.8?
Claude Fable 5 is the first generally available model in Anthropic’s new Mythos class, a tier the company launched on June 9, 2026 to sit above Opus. It scores 80.3% on SWE-Bench Pro versus Opus 4.8’s 69.2%, adds always-on adaptive thinking and automatic context compaction for long agent sessions, and costs exactly double Opus 4.8 per token, both on input and output.
Is Claude Fable 5 more expensive than Claude Opus 4.8?
Yes. Fable 5 costs $10 per million input tokens and $50 per million output tokens, compared to $5 and $25 for Opus 4.8, a precise 2x multiplier confirmed by Anthropic’s documentation and by independent trackers including OpenRouter and Artificial Analysis.
Can developers still use Claude Opus 4.8 now that Fable 5 has launched?
Yes, Opus 4.8 remains fully available through the API and every Claude.ai subscription tier, with no announced deprecation date. Anthropic did not retire it, the model simply moved down to a mid-tier position beneath the new Mythos class, and it remains the cheaper, still-capable option for teams that don’t need Fable 5’s agentic-specific gains.
What is Anthropic’s Mythos class?
Mythos is a new capability tier Anthropic introduced on June 9, 2026, positioned above Opus. Claude Fable 5 is the first publicly released Mythos-class model, built for long-horizon agentic work, with persistent memory, adaptive thinking, and automatic conversation compaction.
Is Claude Mythos 5 the same model as Claude Fable 5?
No. Mythos 5 shares Fable 5’s underlying capabilities but ships without its safety classifiers, and it has reportedly stayed restricted to vetted government and cybersecurity customers rather than becoming a public consumer product.
Does Claude Fable 5 replace Claude Sonnet 5?
No, they serve different budgets. Sonnet 5, which launched June 30, 2026 at $2 and $10 per million tokens, is Anthropic’s cheapest current-generation model and the default for free and standard Pro usage. Fable 5 sits at the opposite end as the premium, agentic-focused tier.
How does Claude Fable 5 compare to Gemini and GPT-5.5?
On coding-specific benchmarks like SWE-Bench Pro, Anthropic’s Fable 5 and Opus 4.8 currently lead independent trackers. GPT-5.5 remains competitive on general-purpose breadth with a 62.9 LLM Stats overall score, while Google’s cheaper Gemini 3.5 Flash tier leads on price-to-performance rather than raw benchmark scores. No single company sweeps every category as of July 2026.
Is Claude Fable 5 included in the free Claude.ai plan?
No. Fable 5 requires a Pro ($20/month), Team ($30/month), or Enterprise subscription, or metered API access. It was briefly available at no extra charge on Pro and Team plans, but that promotional window closed on June 22, 2026.
