Claude 5 is the collective name for Anthropic’s 2026 generation of frontier artificial intelligence models, and it reached the market unlike any Claude generation before it — not as a single launch event, but as three staggered releases across seven weeks. Between 9 June and 24 July 2026, four models arrived under the Claude 5 banner: Claude Fable 5, Claude Mythos 5, Claude Sonnet 5 and Claude Opus 5. Each targets a different budget, workload and risk profile, so choosing between them is now the first decision any Anthropic buyer makes.
The rollout began with Anthropic’s launch announcement for Claude Fable 5 and Claude Mythos 5 on 9 June 2026, continued with Claude Sonnet 5 on 30 June, and closed with Claude Opus 5 on 24 July. Because the family emerged in stages, most machine learning coverage treated each model in isolation. This guide does the opposite: it maps the whole Claude 5 line-up as one purchasing decision, from the price-per-token ladder to the shared platform underneath.
In this Claude 5 guide you will find the full release timeline, the three-rung pricing ladder ($2/$10, $5/$25 and $10/$50 per million tokens), the shared 1M-token context window, the benchmark results together with their caveats, and how each tier behaves inside Claude Code. It sits alongside the other model reviews in our AI models and tools hub, and it complements our broader AI model comparison if you are weighing Anthropic against its rivals.
We also cover the parts of the Claude 5 story that quick summaries gloss over: why Claude Mythos 5 is invitation-only, why Claude Fable 5 traffic carries a 30-day data-retention requirement while Claude Opus 5 carries none, and why there is still no Haiku 5. Every figure in this article comes from Anthropic’s own documentation and announcements, or from named independent coverage listed in the References section.
Table of contents
- What Is the Claude 5 Family?
- Claude 5 Release Timeline: Three Staggered Launches
- The Full Line-Up at a Glance
- Claude 5 Pricing: The Three-Rung Ladder
- Context Windows, Output Limits and Thinking
- Claude 5 Benchmarks: What the Numbers Really Say
- Claude Fable 5 in Depth: The Frontier Flagship
- Claude Mythos 5 and Project Glasswing
- Claude Opus 5 in Depth: The Everyday Frontier Model
- Claude Sonnet 5 in Depth: The Agent Economy Model
- Claude 5 in Claude Code and Agent Harnesses
- Choosing the Right Claude 5 Model for Your Workload
- Upgrading from the Claude 4 Generation
- Frequently Asked Questions
- References
What Is the Claude 5 Family?
According to Anthropic’s model documentation, the Claude 5 family as of August 2026 comprises four models: Claude Fable 5 (API ID claude-fable-5), Claude Mythos 5 (claude-mythos-5), Claude Opus 5 (claude-opus-5) and Claude Sonnet 5 (claude-sonnet-5). Claude Haiku 4.5 remains the current small-tier model, meaning the generation is — for now — a premium and mid-range affair rather than a complete top-to-bottom refresh.
Four Claude 5 Models, Not One
There is no single model literally named “Claude 5”. The phrase describes the generation, the way “Claude 4” loosely described the Opus and Sonnet 4.x releases before it. When someone says they are “using Claude 5”, they usually mean whichever family member their plan or API configuration selects. That ambiguity is worth resolving early, because the four Claude 5 models differ sharply in price, access rules and safety posture.
Why There Is No Haiku 5 in the Claude 5 Line-Up
Anthropic’s documentation still lists Claude Haiku 4.5 as the current small model, with a 200K context window and 64K maximum output tokens. No Haiku 5 exists, and nothing in the published Claude 5 material announces one. If your workloads run on Haiku today, the practical takeaway is simple: the 2026 wave has not yet reached the small tier, so budget-tier planning should assume Haiku 4.5 for now.
A New Tier Above Opus
The biggest structural change in Claude 5 is the appearance of a tier above Opus. Claude Fable 5 and Claude Mythos 5 sit at what Anthropic and its partners call the Mythos class; AWS describes Claude Fable 5 as delivering “Mythos-class capabilities with built-in safeguards”. Opus is no longer the ceiling of the range — it is the everyday frontier tier beneath a research-grade flagship.
Claude 5 Release Timeline: Three Staggered Launches
The staggered schedule is more than trivia: it explains why pricing, defaults and plan placement shifted several times between June and August 2026. Here is the Claude 5 timeline as the primary sources date it.
9 June 2026: Claude Fable 5 and Claude Mythos 5 Arrive
Claude Fable 5 became generally available on 9 June 2026 across the Claude API, Amazon Bedrock, the Claude Platform on AWS, Google Cloud and Microsoft Foundry. The same announcement introduced Claude Mythos 5, the invitation-only variant of the same underlying model. This double launch opened the Claude 5 era at the very top of the capability range rather than in the mainstream.
30 June 2026: Claude Sonnet 5 Goes Default
Three weeks later, Claude Sonnet 5 launched and immediately became the default model for Claude Free and Pro plans. TechCrunch framed it as “a cheaper way to run agents”, noting it could “make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models”. Claude Code received increased rate limits alongside the launch, specifically to support running the model at higher effort levels.
24 July 2026: Claude Opus 5 Completes the Set
Claude Opus 5 was released on 24 July 2026, as reported by Axios. It became the default model on Claude Max, the strongest model available on Claude Pro, and shipped across Claude.ai, Claude Code, Claude Cowork and the API. AWS launched it on Amazon Bedrock the same day, billing it as Anthropic’s most capable Opus model for coding, agents and enterprise workflows. With that, the Claude 5 roster was complete.
August 2026: Claude 5 Pricing Locks In
One postscript matters for anyone budgeting these workloads. Claude Sonnet 5 launched at introductory pricing of $2 per million input tokens and $10 per million output tokens, originally scheduled to rise to $3/$15 after 31 August 2026. In August, Anthropic cancelled the planned increase and made $2/$10 permanent — the Claude Platform docs now list it as the standard price.
The Full Line-Up at a Glance
Before diving into each tier, here is the whole Claude 5 family side by side, using the specifications published in Anthropic’s model documentation.
| Model | Released | API ID | Context | Max output | Access |
|---|---|---|---|---|---|
| Claude Fable 5 | 9 June 2026 | claude-fable-5 | 1M tokens | 128K | API, Bedrock, Google Cloud, Microsoft Foundry |
| Claude Mythos 5 | 9 June 2026 | claude-mythos-5 | Invitation-only | Invitation-only | Project Glasswing partners |
| Claude Sonnet 5 | 30 June 2026 | claude-sonnet-5 | 1M tokens | 128K | Default on Free and Pro, plus API |
| Claude Opus 5 | 24 July 2026 | claude-opus-5 | 1M tokens | 128K | Claude.ai, Claude Code, Claude Cowork, API, Bedrock |
| Claude Haiku 4.5 (carry-over) | Prior generation | — | 200K tokens | 64K | Current small tier |
How the Claude 5 Ladder Fits Together
Read the table bottom-up and the logic of Claude 5 becomes clear. Sonnet 5 is the volume tier — cheap enough to run agents at scale, and to replace the brittle RPA scripts that back-office automation once relied on. Opus 5 is the everyday frontier tier, approaching flagship intelligence at half the flagship’s price. Fable 5 is the peak of what you can buy self-serve, and Mythos 5 is the same intelligence with fewer guardrails, reserved for vetted partners.
Claude 5 Model IDs Across Platforms
Every Claude 5 model ID is a dateless pinned snapshot — claude-fable-5, claude-opus-5, claude-sonnet-5 — a break from the dated snapshot names of earlier generations. On Amazon Bedrock the same IDs take the anthropic. prefix, for example anthropic.claude-opus-5. If your deployment scripts assume date-suffixed IDs, they will need updating before you migrate.
Availability by Surface
Coverage is broad but not uniform. Claude Fable 5 spans the Claude API, Amazon Bedrock, the Claude Platform on AWS, Google Cloud and Microsoft Foundry. Claude Opus 5 runs across Claude.ai, Claude Code, Claude Cowork and the API, with AWS support from day one. Claude Sonnet 5 anchors the Free and Pro plans. Only Claude Mythos 5 sits outside normal channels entirely.
Plan placement is the other half of availability. On the consumer side, Sonnet 5 is what Free and Pro subscribers get by default, Opus 5 is the strongest model a Pro subscriber can reach, and Opus 5 is the default on Max. On the cloud side, the multi-provider spread matters for procurement: teams standardised on AWS, Google Cloud or Microsoft’s stack can each reach the flagship without changing providers.
Claude 5 Pricing: The Three-Rung Ladder
Pricing is where the Claude 5 strategy is easiest to see: each rung doubles or more the rung below it, and each is aimed at a different class of workload. All prices below are per million tokens and come from Anthropic’s announcements and platform documentation.
The chart below visualises the published input prices — $2 for Claude Sonnet 5, $5 for Claude Opus 5 and $10 for Claude Fable 5 per million input tokens (source: Anthropic pricing announcements and the Claude Platform docs).
| Model | Input /M | Output /M | Pricing notes |
|---|---|---|---|
| Claude Sonnet 5 | $2 | $10 | Introductory price made permanent in August 2026; the planned $3/$15 rise was cancelled |
| Claude Opus 5 | $5 | $25 | Same price as Opus 4.8; optional Fast Mode at roughly 2x base price (about $10/$50), API only |
| Claude Fable 5 | $10 | $50 | Less than half the cost of the earlier Claude Mythos Preview; 90% input-token discount with prompt caching |
Claude Sonnet 5: $2/$10, Now Permanent
Claude Sonnet 5 launched at $2 per million input tokens and $10 per million output tokens as an introductory offer through 31 August 2026, with a rise to $3/$15 pencilled in. That rise never happened: Anthropic cancelled it in August and made $2/$10 the standard price. For anyone running high-volume agent fleets, that single decision reshapes the annual budget.
Claude Opus 5: $5/$25 with a Fast Mode
Claude Opus 5 costs $5 per million input tokens and $25 per million output — exactly what Opus 4.8 cost before it — while approaching Claude Fable 5’s frontier intelligence at half the price. The API-only Fast Mode runs roughly 2.5x faster at twice the base price, about $10/$50, which is worth paying when latency, not intelligence, is the bottleneck.
Claude Fable 5: $10/$50 with 90% Prompt Caching
At $10 per million input tokens and $50 per million output, Claude Fable 5 is the most expensive self-serve Claude — yet it arrived at less than half the cost of the earlier Claude Mythos Preview. The 90% input-token discount for prompt caching matters enormously here: cached-heavy agentic workloads pay far less per request than the headline number suggests.
How Claude 5 Costs Compare with Rivals
TechCrunch positioned Claude Sonnet 5 as cheaper than Opus 4.8, GPT-5.5 and Gemini 3.1 Pro at launch, which is precisely the point of the bottom rung: undercut the field for agentic work. Meanwhile Opus 5’s benchmark-per-dollar story — covered below — is how Anthropic argues the middle rung of Claude 5 against everyone else’s flagships.
A Worked Example: What a Big Job Costs
Simple arithmetic on the published prices makes the ladder concrete. A job that reads a full 1M-token context and writes 100K output tokens costs $3 on Claude Sonnet 5 ($2 input plus $1 output), $7.50 on Claude Opus 5 ($5 plus $2.50), and $15 on Claude Fable 5 ($10 plus $5) — before any caching discount.
Prompt caching then changes the flagship’s picture substantially. With Claude Fable 5’s 90% input-token discount applied to cached content, repeated reads of that same 1M-token context bill input at an effective $1 per million rather than $10. Agent loops that revisit the same large context dozens of times per session are exactly where that discount compounds.
Context Windows, Output Limits and Thinking
The Claude 5 generation standardises the platform underneath the price ladder. Whichever tier you buy, the mechanical envelope — context, output and reasoning behaviour — is largely the same, which simplifies switching between tiers mid-project.
The Claude 5 Million-Token Context Window
Claude Fable 5, Claude Opus 5 and Claude Sonnet 5 all carry a 1M-token context window with 128K maximum output tokens. Claude Haiku 4.5, the carried-over small model, remains at 200K context and 64K output. In practice this means every Claude 5 tier can hold an entire large codebase or document archive in a single request — the differentiator between tiers is intelligence and price, not memory.
Output: 128K Standard, 300K in Batches
The 128K maximum output shared by Fable 5, Opus 5 and Sonnet 5 is generous by historical standards, and it changes how you design generation tasks. Whole documentation sets, long structured reports and large code deliverables fit in a single response, which removes an entire class of stitching and continuation logic from application code.
For asynchronous work, the Message Batches API stretches further: Claude Opus 5, Claude Sonnet 5 and the Opus/Sonnet 4.6–4.8 models support up to 300K output tokens via the output-300k-2026-03-24 beta header. Long report generation and bulk transformation jobs that once needed chunking can now emerge in one piece.
Thinking Without budget_tokens
None of the Claude 5 models supports the old extended-thinking budget_tokens mode. Thinking is always on for Claude Fable 5 and adaptive by default on Claude Opus 5 and Claude Sonnet 5. Codebases that manually tuned thinking budgets will need migration — the new lever is the effort parameter, discussed in the Claude Code section below.
Claude 5 Knowledge Cutoffs
Reliable knowledge cutoffs differ across the family: January 2026 for Claude Fable 5 and Claude Sonnet 5, and May 2026 for Claude Opus 5. Counter-intuitively, the mid-tier Opus 5 has the freshest world knowledge in the Claude 5 range, which is worth remembering for research and current-affairs workloads.
Claude 5 Benchmarks: What the Numbers Really Say
Benchmarks sold this generation, so they deserve scrutiny as well as celebration. Below are the headline Claude 5 results from the fact-checked record, including the disputes that vendor pages tend to omit.
Claude Fable 5 on SWE-bench Verified: 95%
Claude Fable 5 scored 95% on SWE-bench Verified, a figure confirmed on the independent vals.ai leaderboard. That independent confirmation matters: it is the cleanest single number in the entire Claude 5 benchmark story, and it established the flagship’s software-engineering credentials on day one.
SWE-Bench Pro: 80.3% — and a Dispute
On the harder SWE-Bench Pro, Claude Fable 5 posted 80.3% against Claude Opus 4.8’s 69.2% — roughly 11 points ahead of the field — though the figure was produced with Anthropic’s own scaffolding. Independent evaluators contest the headline: three sources report three different SWE-Bench Pro leaders depending on the harness used. Treat the gap as directionally real but harness-sensitive.
The chart below shows the two figures just quoted — 80.3% for Claude Fable 5 versus 69.2% for Claude Opus 4.8 on SWE-Bench Pro, per Anthropic’s scaffolded run.
FrontierCode and the Hebbia Finance Benchmark
Beyond SWE-bench, Claude Fable 5 achieved the highest score among frontier models on Cognition’s FrontierCode benchmark at medium effort, plus the highest score for senior-level reasoning on the Hebbia Finance Benchmark. The finance result is notable because it extends the Claude 5 case beyond coding into white-collar analytical work.
Claude Opus 5’s Record Run
Claude Opus 5 set new state-of-the-art results on Frontier-Bench v0.1 (double Opus 4.8’s score), ARC-AGI 3 (three times the next-best model) and Zapier AutomationBench (roughly 1.5x the next-best pass rate). Its value story is sharper still: on CursorBench 3.2 it lands within 0.5% of Claude Fable 5’s peak at half the cost, and on OSWorld 2.0 it actually surpasses Fable 5 at one-third the cost.
Claude Sonnet 5’s Agentic Coding Numbers
On agentic coding, Claude Sonnet 5 scored 63.2% against Claude Opus 4.8’s 69.2% and Claude Sonnet 4.6’s 58.1%, while slightly outperforming Opus 4.8 on knowledge work. As the chart below shows, the cheapest tier of the generation lands within about six points of the previous Opus flagship on agentic coding, at a fraction of its price (figures per TechCrunch).
Claude Fable 5 in Depth: The Frontier Flagship
Claude Fable 5 is the model the rest of the Claude 5 family is measured against — the self-serve flagship that AWS bills as Mythos-class capability with the safety rails still attached. Four aspects define it in practice.
Vision: From Screenshot to Working App
In vision testing, Claude Fable 5 rebuilt working web applications from screenshots alone and completed Pokémon FireRed without navigation aids. For product teams, the screenshot-to-app result is the practical one: it signals that visual reverse-engineering of interfaces is now a dependable Claude 5 workflow rather than a demo trick.
The Pokémon result is easy to dismiss as a stunt, but it demonstrates something specific: sustained visual navigation of an unfamiliar environment without hand-built aids. That is the same competence an agent needs to operate real software through a screen — reading state, planning a route and recovering from wrong turns over a long horizon.
Science: Protein Design at Roughly 10x Speed
In drug-design evaluations, Claude Fable 5 accelerated protein design roughly 10x, with 9 of 14 targets yielding strong drug candidates. In blinded molecular-biology comparisons it was preferred about 80% of the time over Opus-class models. This is the workload class — long-horizon scientific reasoning — that justifies the flagship’s premium over the rest of the Claude 5 range.
Three Built-In Safeguard Classifiers
Claude Fable 5 ships with three safeguard classifiers: a cybersecurity classifier that blocks offensive cyber tasks, a biology/chemistry classifier that falls back to Claude Opus 4.8, and an anti-distillation classifier. These are not optional settings — they are the “built-in safeguards” half of the Mythos-class equation, and they shape what the model will and will not do.
The 30-Day Data-Retention Rule
Mythos-class traffic carries a 30-day data-retention requirement, with no use of that data for model training. Compliance teams should note the asymmetry inside Claude 5: Claude Opus 5 carries no special retention requirement, so organisations with strict data-residency rules may find the mid tier easier to approve than the flagship.
Claude Mythos 5 and Project Glasswing
Every Claude 5 explainer mentions Mythos 5; few explain it plainly. It is not a bigger model — it is the same one, with different rules attached.
The split-release design answers a real problem. Frontier capability in cybersecurity and biology is dual-use: the same skills that harden systems or design proteins can be misused. Anthropic’s answer in this generation was to ship one set of weights twice — once with classifiers for the public, once without them for vetted specialists — rather than to hold the capability back entirely.
One Model, Two Names
Claude Mythos 5 is the same underlying model as Claude Fable 5, with safeguards lifted for authorised users. Where Fable 5’s classifiers block offensive cyber work and route sensitive biology to a fallback, Mythos 5 removes those interventions for vetted purposes. Same weights, different guardrails — that is the entire difference.
Who Can Access Claude Mythos 5
Access is restricted to Project Glasswing cybersecurity partners and select biology researchers. It is invitation-only, with no self-serve sign-up of any kind. For virtually every commercial buyer reading a Claude 5 comparison, Mythos 5 is therefore a non-option — relevant to understanding the family’s architecture, irrelevant to the purchase decision.
What Mythos-Class Means for Claude 5 Buyers
The practical meaning of “Mythos-class” for ordinary buyers is Claude Fable 5: the full capability of the class, delivered behind classifiers. Anthropic’s own materials note that Claude Opus 5 remains behind Mythos 5 on cybersecurity and biology research capability, which confirms the ceiling of the family sits with the Fable/Mythos pair, not with Opus.
Claude Opus 5 in Depth: The Everyday Frontier Model
If Fable 5 is the statement piece, Claude Opus 5 is the model most Claude 5 buyers will actually standardise on. Released last, it landed with the best price-to-capability ratio in the family.
The Effort Dial: Low Through Max
Claude Opus 5 features an effort dial running from low through max, letting users trade computing power against cost. Fortune highlighted the design at launch: at lower effort levels the model preserves much of its performance while using fewer tokens. In effect, one model spans several price-performance points that previously required switching between models.
The dial also changes how teams should think about model selection. Instead of maintaining separate configurations for cheap and expensive paths, a single Claude Opus 5 deployment can serve both: routine requests run at lower effort, and the hard ones escalate to max without a model swap, a redeploy or a second integration.
The Most Aligned Claude Model to Date
Claude Opus 5 recorded a 2.3 misalignment score on Anthropic’s behavioural audit — the most aligned Claude model to date. For enterprises weighing agent autonomy, that number is arguably as important as any capability benchmark in this guide: the generation being trusted with days-long tasks is also the one auditing best on behaviour.
Alignment does not mean unlimited capability, though. Anthropic notes that Opus 5 remains behind Mythos 5 on cybersecurity and biology research capability — a deliberate gap. The mid tier is tuned to be the safest broadly available option, not the most powerful one in absolute terms.
Fast Mode on the API
The API-only Fast Mode runs roughly 2.5x faster at twice the base price — about $10/$50 per million tokens. Interactive products where response latency drives user experience can toggle it per workload, paying Fable-level prices for Opus-level output at materially higher speed.
No Special Data-Retention Requirement
Unlike Claude Fable 5 with its 30-day Mythos-class retention rule, Claude Opus 5 carries no special data-retention requirement. Combined with its default placement on Claude Max and top billing on Claude Pro, that makes it the path of least resistance for regulated organisations adopting Claude 5.
Claude Sonnet 5 in Depth: The Agent Economy Model
Claude Sonnet 5 is the volume play — the tier designed to make Claude 5 agents cheap enough to run everywhere, all the time.
Default for Free and Pro
Since 30 June 2026, Claude Sonnet 5 has been the default model for Claude Free and Pro plans, which makes it the Claude 5 model most people have actually used. Its launch positioning was explicit: agentic capability that recently required larger, costlier models, now at $2/$10 — pricing that August 2026 made permanent.
Fewer Hallucinations, Less Sycophancy
Anthropic reports that Claude Sonnet 5 shows lower rates of hallucination and sycophancy than Claude Sonnet 4.6. For agent builders this is the quiet headline: an agent that flatters its operator or invents tool output is worse than useless, so behavioural improvements compound across every step of a long-running pipeline.
Consider what an agent actually does with a hallucination: it acts on it. A fabricated file path becomes a failed command; an invented API response becomes a corrupted workflow. Reliability gains at the model level therefore translate directly into fewer human interventions per task — which is the real cost metric for automation at scale.
Deliberately Limited Offensive Cyber
Claude Sonnet 5 scored 0% full success on Firefox 147 exploit tests — a deliberate design choice reflecting reduced offensive-cyber capability versus Opus-tier models. The family’s safety architecture is graduated: the cheaper and more widely available the tier, the harder its ceiling on dangerous capability.
Seen across the whole range, the pattern is consistent. Sonnet 5 is capability-limited by design, Fable 5 is capability-rich but classifier-guarded, and Mythos 5 is unguarded but access-controlled. Buyers rarely notice this architecture until a workload trips a boundary, so it is worth knowing which tier draws its lines where before you commit.
Claude 5 in Claude Code and Agent Harnesses
The Claude 5 generation was clearly built with agent harnesses in mind, and Claude Code is the reference harness. Three operational details determine what you get out of it.
Effort Defaults: High by Default, xhigh by Choice
The effort parameter defaults to high for Claude Opus 5 and Claude Sonnet 5 on both the Claude API and Claude Code. Anthropic’s guidance is to opt in to xhigh deliberately for coding and agentic work rather than leave it on for everything. Since none of the Claude 5 models supports the old budget_tokens thinking mode, effort is now the primary quality-versus-cost lever.
In day-to-day use, the sensible pattern follows the guidance directly: leave the default in place for exploratory sessions and code review, and switch to xhigh when handing the harness a well-specified, high-stakes implementation task. Because effort trades tokens for quality, an xhigh session on a vague brief mostly buys expensive wandering.
Rate Limits Raised for Claude 5
Claude Code received increased rate limits alongside the Sonnet 5 launch, precisely so users could run the model at higher effort levels without throttling. If you benchmarked Claude Code before 30 June 2026 and found limits constraining, that finding is stale — re-test against the current allowances.
Agents That Run for Days
Run in an agent harness such as Claude Code or Claude Managed Agents, Claude Fable 5 can work for days at a time — planning across stages, delegating to sub-agents and checking its own work. That is the operational definition of the Claude 5 flagship tier: not a chat model that codes, but a system that sustains multi-day engineering campaigns.
Choosing the Right Claude 5 Model for Your Workload
With the specifications, prices and benchmarks on the table, the decision reduces to a handful of questions. Our team builds this kind of selection into client AI strategy engagements, and the same logic applies whether you deploy one Claude 5 model or a routed mix.
| Workload | Best Claude 5 pick | Why |
|---|---|---|
| High-volume agents and tools | Claude Sonnet 5 | $2/$10 permanent pricing; 63.2% agentic coding within six points of Opus 4.8 |
| Everyday frontier coding and reasoning | Claude Opus 5 | Within 0.5% of Fable 5 on CursorBench 3.2 at half the cost; effort dial for budget control |
| Latency-sensitive products | Claude Opus 5 Fast Mode | Roughly 2.5x faster at about $10/$50, API only |
| Days-long autonomous engineering, science | Claude Fable 5 | 95% SWE-bench Verified; multi-day agent runs; 90% prompt-caching discount |
| Strict data-retention constraints | Claude Opus 5 | No special retention rule, unlike Fable 5’s 30-day Mythos-class requirement |
| Small, cheap classification tasks | Claude Haiku 4.5 | No Haiku 5 exists; the 4.5 small tier remains current |
Which Claude 5 Model for Coding?
For interactive coding, Claude Opus 5 is the default answer: near-flagship benchmark results at half the flagship price, with the effort dial to trim costs on routine tasks. Reserve Claude Fable 5 for the hardest problems and multi-day autonomous refactors, where its 95% SWE-bench Verified score and days-long harness endurance earn the premium.
Which Claude 5 Model for Agents?
Start agent fleets on Claude Sonnet 5. Its economics are the whole point — permanent $2/$10 pricing against capability that sits within striking distance of the previous Opus flagship. Escalate individual agent roles to Claude Opus 5 where reliability or reasoning depth measurably pays, a routing pattern we implement in intelligent automation projects.
Which Claude 5 Model for Regulated Data?
Claude Opus 5, in most cases. The Claude 5 family splits on data handling: Fable 5 traffic carries the 30-day Mythos-class retention requirement (with no training use), while Opus 5 carries no special requirement at all. Where legal review is the bottleneck, the mid tier clears it faster.
When Claude Haiku 4.5 Is Still Enough
Not every task needs a Claude 5 model. Simple extraction, classification and routing jobs that fit comfortably in 200K context and 64K output still run happily on Claude Haiku 4.5 — and until a Haiku 5 appears, that carried-over small model remains the family’s official budget answer.
Building a Claude 5 Routing Strategy
Mature deployments rarely pick one tier and stop. The pattern that works is a router: Sonnet 5 handles the volume, Opus 5 takes anything the router flags as hard or compliance-sensitive, and Fable 5 is reserved for the long-horizon jobs that justify flagship rates. Because all three share the 1M-token context and the same API shape, escalation between tiers requires no re-architecture — just a different model ID on the request.
Upgrading from the Claude 4 Generation
Most teams arrive at Claude 5 from an existing Claude 4.x deployment, so the migration story matters as much as the launch story.
What Changed Since Opus 4.7 and 4.8
We reviewed the previous generation in depth in our Claude Opus 4.7 guide, and the Claude 5 deltas against its successors are stark: Claude Opus 5 doubled Opus 4.8’s Frontier-Bench v0.1 score at identical $5/$25 pricing, and Claude Fable 5 beat Opus 4.8 by roughly 11 points on SWE-Bench Pro. Same money, materially more model.
Thinking Mode Migration
The breaking change is extended thinking. The old budget_tokens mode is unsupported across Claude 5: thinking runs always-on for Claude Fable 5 and adaptively by default on Claude Opus 5 and Claude Sonnet 5. Replace thinking-budget logic with effort selection, and remember model IDs are now dateless pinned snapshots rather than dated releases.
The 4.6–4.8 Models Still Matter
The previous generation has not vanished from the platform. The 300K-output Batches beta covers Opus and Sonnet 4.6–4.8 alongside the new models, and Claude Fable 5’s biology/chemistry classifier literally falls back to Claude Opus 4.8. Claude 5 supersedes the 4.x line commercially, but operationally the two generations still interlock.
A Practical Migration Order
For most teams the low-risk sequence is: move consumer-plan usage first (it has already happened by default), migrate API agent fleets from Sonnet 4.6 to Sonnet 5 next for the pricing and behaviour gains, then re-point Opus 4.8 workloads at Claude Opus 5 — same $5/$25 price, materially better results. Leave the flagship evaluation until last, when you have real usage data to judge whether Fable-tier jobs exist in your stack at all.
Frequently Asked Questions
Is there a single model called Claude 5?
No. Claude 5 is the generation name covering four models: Claude Fable 5, Claude Mythos 5, Claude Opus 5 and Claude Sonnet 5. There is no API ID claude-5; you always select a specific family member such as claude-opus-5 or claude-sonnet-5.
Does a Claude Haiku 5 exist?
No. Anthropic’s model documentation lists Claude Haiku 4.5 as the current small-tier model, with 200K context and 64K output. Nothing announced in the June–July 2026 releases includes a small-tier refresh.
What does Claude 5 cost right now?
Per million tokens: Claude Sonnet 5 costs $2 input/$10 output (made permanent in August 2026), Claude Opus 5 costs $5/$25 with an optional API-only Fast Mode at about $10/$50, and Claude Fable 5 costs $10/$50 with a 90% input discount for prompt caching.
Can anyone use Claude Mythos 5?
No. Claude Mythos 5 — the same model as Fable 5 with safeguards lifted — is restricted to Project Glasswing cybersecurity partners and select biology researchers. It is invitation-only, with no self-serve sign-up.
Do Claude 5 models still support thinking budgets?
No. The extended-thinking budget_tokens mode is unsupported across the family. Thinking is always on for Claude Fable 5 and adaptive by default on Claude Opus 5 and Claude Sonnet 5, with the effort parameter as the new control.
How does Claude Code run Claude 5 models?
Claude Code defaults the effort parameter to high for Claude Opus 5 and Claude Sonnet 5, with Anthropic recommending a deliberate opt-in to xhigh for demanding coding and agentic work. Rate limits were raised at the Sonnet 5 launch, and Claude Fable 5 can sustain multi-day agent runs inside the harness.
Which model has the freshest knowledge?
Claude Opus 5, with a reliable knowledge cutoff of May 2026. Claude Fable 5 and Claude Sonnet 5 both sit at January 2026. If your workload depends on recent-world knowledge rather than raw reasoning power, the mid tier is — unusually — the best-informed choice.
Are the models available on Amazon Bedrock?
Yes. Claude Fable 5 was available on Amazon Bedrock from its 9 June launch, and AWS launched Claude Opus 5 on Bedrock at its release on 24 July. On Bedrock the IDs carry the anthropic. prefix — for example anthropic.claude-opus-5.
References
Claude Fable 5 and Claude Mythos 5 — Anthropic
Introducing Claude Opus 5 — Anthropic
Introducing Claude Sonnet 5 — Anthropic
Models overview — Claude Platform Docs
Anthropic launches Claude Sonnet 5 as a cheaper way to run agents — TechCrunch
Anthropic releases new model, Opus 5 — Axios
Anthropic debuts Claude Opus 5 with cost-capability toggle — Fortune
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
