Every Claude Model: Full Guide from Claude 3 to Opus 5.5

Anthropic has released over a dozen Claude models since March 2024. Each generation brought meaningful capability jumps, pricing shifts, or entirely new features. This page tracks every model in one place so you can compare specs, understand the evolution, and pick the right model for your work.

Complete Model Timeline

Which Model Should You Use Today?

For most developers using Claude Code, the choice comes down to four active models:

Daily driver: Sonnet 5. The most agentic Sonnet yet and the new default on Free and Pro, which supersedes Sonnet 4.6. It scores 63.2% on SWE-bench Pro (closing in on Opus 4.8's 69.2%), edges past Opus 4.8 on knowledge work, and drives terminals and browsers autonomously. It costs $2/$10 per million tokens: Anthropic made the launch price standard and cancelled the scheduled increase to $3/$15. It handles 90%+ of coding tasks without compromise.

Heavy lifting and frontier work: Opus 5.5. Anthropic's recommended starting point for most workloads and the default Opus model in Claude Code, which supersedes Opus 5. It leads Opus 5 and Fable 5.1 on all nine benchmarks in its launch table, including 66.4% on Terminal-Bench 4.0 against Fable 5.1's 55.8%, at $4/$20 per million tokens with cache reads at $0.20, and Anthropic estimates it runs about 40% cheaper than Opus 5 on typical workloads. Its default effort is medium, and on Terminal-Bench 4.0 that default beats Opus 5 at max for about a fifth of the cost. Use it for complex architecture decisions, long agentic sessions, browser and computer use, and agent teams. Four API changes break Opus 5 code, including thinking that cannot be disabled, so read the Opus 5.5 breaking changes before switching.

The tier above, when Opus 5.5 at higher effort falls short: Fable 5.1. The Fable 5 successor shipped September 1, 2026 at the same $10/$50 list price with cache reads cut 75% to $0.25 per million tokens, which Anthropic estimates makes typical Fable workloads about 25% cheaper and highly agentic ones up to about 45% cheaper. Opus 5.5 now outscores it on every row of the September 22 launch table, and Anthropic says the real-world gap is narrower than those scores suggest, so Fable 5.1 is the escalation for demanding reasoning and long-horizon agentic work where your own evals show Opus 5.5 falling short. Plan terms carry over from Fable 5: included on Max plans and premium seats up to 50% of weekly usage limits, with Pro and standard seats on usage credits. Three API changes break Fable 5 code, so read the breaking changes before switching.

Budget tasks: Haiku 4.5. Smart model switching routes simple tasks here automatically. Good for file reads, quick edits, and routine questions. The cheapest option that still delivers quality output.

For a deeper dive into tactical model switching during development sessions, see the model selection guide. For managing costs across your team, check the usage optimization guide. If you want these routing strategies applied out of the box, ClaudeFast's Code Kit ships with model routing configuration for all active Claude models.

How Pricing Evolved

One of the most important trends across Claude generations is falling costs. The original Claude 3 Opus cost $15/$75 per million tokens. Opus 4.5 through Opus 5 held the standard tier at $5/$25, a 67% reduction, and Opus 5.5 became the first Opus to cut it further, to $4/$20, while taking the top of the benchmark table. Meanwhile, capabilities have increased dramatically.

Opus 4.8 introduced a Fast mode at $10 input / $50 output per 1M tokens that runs at roughly 2.5x the speed of the standard tier, and Opus 5 carries it forward at the same rate as a research preview on the Claude API. Fable 5.1 took a different route to the same end in September 2026: the $10/$50 list price stayed, but cache reads fell 75% to $0.25 per million tokens, so a Fable bill dominated by re-read context drops by up to about 45% with no price cut on the rate card. Opus 5.5 three weeks later did both: base rates fell 20% to $4/$20, cache reads fell 60% to $0.20, and Fast mode dropped to $8/$40. The trend is clear: each generation delivers more capability per dollar spent, whether through a lower list price or a cheaper token mix.

Key Milestones

March 2024: The three-tier model with Claude 3. Anthropic established the Opus/Sonnet/Haiku naming convention and introduced vision capabilities. This was the foundation everything else built on.

June 2024: Mid-tier disruption with Claude 3.5 Sonnet. A cheaper model outperforming the flagship changed how developers thought about model selection. Bigger stopped meaning better.

October 2024: Computer Use with Claude 3.5 Sonnet v2. The first major AI model to directly control a computer interface. Cursor movement, clicking, typing. A new category of agentic capability.

February 2025: Extended thinking with Claude 3.7 Sonnet. Hybrid reasoning let Claude pause and think step-by-step before responding. Quality jumped on complex math, science, and multi-step code problems.

May 2025: Claude 4 generation with Claude 4. Professional-grade coding capabilities made Claude Code a daily driver for serious development work.

November 2025: The efficiency revolution with Opus 4.5. 67% price cut and 76% fewer output tokens made premium intelligence affordable for everyday use.

February 2026: 1M context and Agent Teams with Opus 4.6. The largest context window (now GA with unified pricing as of March 2026) and native multi-agent collaboration open new categories of work that weren't possible before.

February 2026: Sonnet catches Opus with Sonnet 4.6. For the first time, a Sonnet model is preferred over the previous generation's Opus in coding evaluations. Computer use accuracy hit 94% on insurance benchmarks.

April 2026: Vision and verification with Opus 4.7. A 3x jump in image resolution (2,576px, 3.75 megapixels) makes vision genuinely reliable for professional use. The model verifies its own outputs, follows instructions more literally, and introduces the xhigh effort level. CursorBench jumped from 58% to 70%.

April 2026: The frontier crosses a line with Mythos Preview. Anthropic's most capable model found thousands of zero-day vulnerabilities autonomously, including 27-year-old OpenBSD bugs. Deemed too powerful for general release, Mythos ships only through Project Glasswing to critical-infrastructure partners at $25/$125 per million tokens. Opus 4.7 becomes the deliberately narrower test vehicle for Mythos-era safeguards.

May 2026: Reliability and agentic browser with Opus 4.8. The first Claude model where browser agents land squarely in production territory (84% on Online-Mind2Web) and the first where Anthropic quantifies a reliability gain rather than a benchmark gain: roughly four times less likely than Opus 4.7 to let flaws in its own code pass unremarked. A new Fast mode runs at 2.5x the speed of standard for $10/$50 per million tokens, three times cheaper than previous Fast tiers, and Dynamic Workflows in Claude Code lets the same model coordinate hundreds of parallel subagents with adversarial verification.

June 2026: The frontier goes public with Fable 5 and Mythos 5. Anthropic took the Mythos-class capability it gated behind Project Glasswing in April and shipped a safeguarded version, Fable 5, to everyone. A classifier system routes high-risk cybersecurity, biology, chemistry, and distillation requests to Opus 4.8 instead, triggering in under 5% of sessions. Mythos 5, the same model with cyber safeguards lifted, stays restricted to Glasswing partners. Pricing doubles to $10/$50 per million tokens. Subscription access then took a detour: a US export-control directive disabled both models worldwide on June 12, access was restored on July 1 after the controls lifted, and the plan terms settled on July 20 with Fable 5 included on Max and premium seats and metered by usage credits on Pro and standard seats.

June 2026: Sonnet closes on Opus with Sonnet 5. Anthropic's most agentic Sonnet lands as the default model on Free and Pro, scoring 63.2% on SWE-bench Pro against Opus 4.8's 69.2% and slightly outperforming Opus 4.8 on knowledge work. It plans, drives browsers and terminals, and finishes long multi-step tasks autonomously, launching at $2/$10 per million tokens, a price Anthropic later made standard instead of raising it to $3/$15. Sonnet 5 supersedes Sonnet 4.6 and makes near-flagship agentic coding a daily-driver expense.

July 2026: The price line stops moving with Opus 5. For the first time a Claude release closes most of the gap to the frontier tier without a price increase: Opus 5 holds $5/$25, more than doubles Opus 4.8 on Frontier-Bench (43.3 vs 18.9), scores 30.2 on ARC-AGI-3 where Opus 4.8 managed 1.5, and wins seven of the eight quantified evals against Fable 5 at half its rate. It ships with thinking on by default, an explicit max effort tier, a 512-token prompt-cache minimum, and the lowest misaligned-behavior score of any Claude model. Opus 5 becomes the default on Claude Max and supersedes Opus 4.8, which stays available as the classifier fallback.

August 2026: Claude output becomes provable with text watermarking. To comply with the EU AI Act, Anthropic began keying Claude's word choices to a cryptographic key, making outputs statistically detectable with no hidden characters, no extra tokens, and no added cost. The watermark carries no user-identifying information, applies globally, and a detection API is on the way.

September 2026: The Fable bill drops without a price cut with Fable 5.1 and Mythos 5.1. Fable 5.1 keeps Fable 5's $10/$50 list price, cuts cache reads 75% to $0.25 per million tokens, and leads Fable 5 and Opus 5 on all seven published benchmarks (and GPT-5.6 Sol wherever Sol is charted), more than doubling Fable 5 on Terminal-Bench-Science (52.6 vs 24.7). At low effort it beats Fable 5 at max for about a quarter of the cost per task. Three API changes break Fable 5 code: forced tool use returns a 400, older models cannot read its thinking blocks, and editing earlier turns invalidates thinking blocks on API accounts created from August 31, 2026, an anti-distillation measure. Cyber safeguards intervene about 60% less per Claude Code session and now allow vulnerability identification, biology safeguards fire 85% less on benign questions, and Enterprise Frontier Safeguards will bring zero-retention privacy to the Fable line from fall 2026. Mythos 5.1, the same model with permissive safeguards, is limited to vetted cyber and life-sciences organizations in the US. Fable 5 and Mythos 5 move to legacy status.

September 2026: Opus overtakes the Fable tier with Opus 5.5. The first model in the Claude 5.5 family leads Opus 5 and Fable 5.1 on all nine benchmarks in its launch table, from 66.4% on Terminal-Bench 4.0 to 1,846 Elo on GDPval-AA v2.1, while cutting the Opus price to $4/$20 and cache reads to $0.20, about 40% cheaper to run than Opus 5 by Anthropic's estimate. Its default effort drops to medium, and on Terminal-Bench 4.0 that default beats Opus 5 at max for about a fifth of the cost. It inherits Fable 5.1's API restrictions (always-on thinking, no forced tool use, thinking blocks bound to the conversation), adds Fable-class biology safeguards, and scored better than any recent Claude model on nearly every measure of misaligned behavior. Anthropic now names it the starting point for most workloads, and Opus 5 moves to legacy status.

Each model page in this section covers the full specs, benchmarks, pricing, and practical guidance for that specific release. Click any model in the timeline above to get the details.