What's the Next Claude Model? Anthropic's New Roadmap

To work out what comes next, you have to see the shape of what came before. The Claude lineage is one of the clearest cadence stories in the industry, starting slow and then accelerating hard through 2025 and 2026.

Here is the confirmed timeline, drawn from Anthropic's own announcements and contemporaneous reporting.

• March 2023: Claude 1, Anthropic's first public model.
• July 2023: Claude 2, the first model most people could try freely.
• November 2023: Claude 2.1, with a longer context window and fewer refusals.
• March 2024: The Claude 3 family lands as a trio, Haiku, Sonnet and Opus, the small, mid and large tiers that still define the lineup.
• June 2024: Claude 3.5 Sonnet, a mid-tier model that punched above the original Opus.
• October to November 2024: Claude 3.5 Sonnet v2 and Claude 3.5 Haiku.
• February 2025: Claude 3.7 Sonnet, the first hybrid reasoning model.
• May 2025: Claude Opus 4 and Sonnet 4, the start of the version 4 generation.
• August 2025: Claude Opus 4.1.
• September to October 2025: Claude Sonnet 4.5 and Haiku 4.5.
• November 2025: Claude Opus 4.5.
• February 2026: Claude Opus 4.6 and Sonnet 4.6 arrive close together.
• April 2026: Claude Opus 4.7.
• 28 May 2026: Claude Opus 4.8, the current general-purpose flagship.
• 9 June 2026: The Mythos-class arrives publicly with Fable 5 (safeguarded) and Mythos 5 (unrestricted, partner-only).
• 12 June 2026: Fable 5 and Mythos 5 suspended worldwide under a US export-control directive.
• 30 June 2026: Claude Sonnet 5 launches, delivering near-Opus performance at a lower price and finally resolving months of Sonnet 5 speculation.
• 1 July 2026: Fable 5 and Mythos 5 access is restored, ending the 19-day suspension.
• 9 July 2026: A research checkpoint codenamed "Claude Honeycomb EAP" briefly appears in Cursor's model picker, widely read as an early Opus 5 preview, then disappears within hours.
• Late July 2026: Opus 5 leak cycle intensifies, with multiple outlets citing 23 July as a likely launch date. As of 24 July, Anthropic has not confirmed a release.

There is one important strand that sits alongside the numbered models: the Mythos-class. Before the public Mythos 5, Anthropic ran a Mythos Preview, an unreleased frontier model with extreme coding and vulnerability-finding ability, kept out of public hands and instead fed into Project Glasswing, a cyber-defence initiative. Mythos 5 was the productised version of that lineage, and Fable 5 was its safeguarded public sibling built on the same weights.

Two naming notes worth fixing, because the rumour mill keeps getting them wrong. There was never a Claude Sonnet 4.7, the line went 4.6 then jumped to a planned 4.8. And the leaked early-2026 checkpoint codenamed Fennec, which half the internet assumed was Sonnet 5, shipped as the far more modest Sonnet 4.6.

Stepping back, the lineage tells a story in three acts. Act one, from Claude 1 in 2023 to the Claude 3 family in 2024, was about establishing the three-tier Haiku, Sonnet and Opus structure and proving the models could compete at the frontier. Act two, through 2025, was about reasoning: Claude 3.7 introduced hybrid thinking, and the version 4 line turned that into a reliable, agentic workhorse. Act three, in 2026, is about speed and specialisation: a blistering point-release cadence on Opus and Sonnet, plus the entirely new Mythos-class aimed at the very top of capability. The next model belongs to act three, which is why cadence and the Mythos-class track matter more to the prediction than any single benchmark.

It is also striking how the centre of gravity has shifted within the lineup. In 2024, Sonnet was the headline-grabber, the mid-tier model that outran the previous Opus. In 2026, Opus reclaimed the spotlight as the fast-moving flagship, Sonnet fell a step behind at 4.6, and the genuine frontier conversation moved up again to the Mythos-class. That drift upward is part of why a Sonnet refresh feels overdue, and why the most exciting "next" candidates sit at the Mythos-class top of the range rather than in the middle tiers.

The single most useful signal for predicting the next Claude model is the gap between the last few. The pattern is unambiguous: it is shrinking fast.

In the early years the gaps were measured in quarters. Claude 1 to Claude 2 was around four months. Claude 2 to the Claude 3 family was about eight months. Claude 3 to Claude 3.5 Sonnet was roughly three months. Through 2025 the Opus line moved from version 4 in May to 4.1 in August to 4.5 in November, broadly a quarterly rhythm.

Then 2026 broke the pattern. Look at the Opus cadence alone:

• Opus 4.6: early February 2026.
• Opus 4.7: mid-April 2026, roughly ten weeks later.
• Opus 4.8: late May 2026, roughly six weeks after that.
• Fable 5 (Mythos-class): 9 June 2026, less than two weeks after Opus 4.8.

That is a clear acceleration, from quarterly to roughly six-to-eight-week point releases, with the Mythos-class layered on top as a separate track. The implication is straightforward. If nothing had disrupted the rhythm, the next new Claude model would have been due in late June or July 2026, on pure trend extrapolation.

But something did disrupt it. The 12 June recall means Anthropic's near-term engineering attention is split between shipping the next thing and getting its flagship Mythos-class models back online. That tension is the heart of the whole roadmap question right now.

It is worth pausing on why the cadence accelerated at all, because the reasons tell you whether it can continue. Three forces compressed the gaps. The first is method maturity: Anthropic moved to a hybrid reasoning architecture with Claude 3.7 in early 2025, and once that base was in place, improvements arrived as incremental post-training and tuning passes rather than full rebuilds. Point releases like 4.6, 4.7 and 4.8 are exactly the kind of thing you can ship every few weeks when the underlying recipe is stable. The second is competitive pressure, which we cover later. The third is sheer capacity, the compute that lets you train and evaluate candidates in parallel rather than in series.

There is also a naming pattern hiding in the cadence that helps decode what "next" looks like. Anthropic uses the decimal to signal scope. A jump like 4.5 to 4.6 tends to mean a broad refresh across capabilities. A smaller-feeling step like 4.7 to 4.8 has tended to mean targeted gains, in the Opus 4.8 case on agentic reliability, long-horizon task completion and calibration rather than a headline reasoning leap. If the next release is another decimal bump, expect it to be narrow and fast. If Anthropic ever does reach for a whole number, that is the signal to expect a genuine architecture or training-scale change, and so far it has consistently declined to send that signal.

One more cadence nuance matters for prediction. The Mythos-class is on its own clock, decoupled from the Opus and Sonnet point-release rhythm. Mythos Preview ran quietly inside Project Glasswing for weeks before any public model appeared, then Fable 5 and Mythos 5 arrived as a pair. That means the next Mythos-class move does not have to wait for, or align with, the next Opus or Sonnet number. The two tracks can and do move independently, which is part of why pinning a single "next Claude model" date is so hard.

You cannot answer "what's next?" without understanding what just happened. On 12 June 2026, three days after launch, the US government ordered Anthropic to suspend Fable 5 and Mythos 5 for all foreign nationals under an export-control directive citing national security. Rather than try to fence access by nationality across a global product, Anthropic switched both models off for everyone worldwide.

The trigger, as Anthropic described it, was a narrow cyber jailbreak: getting the model to read a codebase and fix software flaws, a dual-use capability that is exactly what defenders do daily and exactly what an attacker would want. Anthropic complied with the directive but publicly disagreed, arguing the capability is widely available from other models including OpenAI's GPT-5.5, and that applying this standard industry-wide "would essentially halt all new model deployments." We covered the full episode in Claude Fable 5 Banned: Why It Was Pulled Offline.

The recall reshapes the roadmap in three ways. First, it adds a regulatory gate that did not exist before: any future Mythos-class release now has to clear an export-control review that can move in hours. Second, it makes the safeguarded versus unrestricted split central to product strategy, because Fable, the safe one, was the one that got pulled. Third, it puts a question mark over the exact thing the cadence pointed to, a fast follow-on frontier model.

On restoration, Anthropic said it was working to bring access back as soon as possible. Our detailed read on the timing, the three routes back, and the identity-verification rollout that could matter is in When Is Claude Fable 5 Coming Back?. The short version: there is no official date, restoration is more likely than not in the near term, and verified US access could plausibly return before full international access.

The deeper significance for the roadmap is the precedent. Before 12 June, the binding constraints on a new Claude model were internal: did it pass Anthropic's own red-team and Responsible Scaling Policy thresholds, and was it commercially ready. After 12 June, there is a fourth, external constraint that can fire after launch, not just before it. A model can pass every internal gate, ship, and still be pulled by directive within days. No amount of pre-release safety work fully removes that risk, which changes how a lab thinks about when, and to whom, it exposes its most capable systems.

That precedent cuts directly against the unrestricted half of the Mythos-class strategy. Mythos 5 was never a public product; it was the unsafeguarded sibling reserved for vetted Project Glasswing partners precisely because of its cyber-offence ability. Yet it was named in the same suspension as Fable. If even the carefully gated partner model attracts an export-control order, Anthropic's calculus for the next Mythos-class release has to weight regulatory exposure far more heavily than capability marketing. Expect the next high-capability release to come wrapped in more conspicuous safeguards, clearer access tiers, and possibly geographic or identity gating from day one rather than as an afterthought.

There is a counter-pressure too, and it is commercial. Fable 5 was pitched as Anthropic's most capable widely released model, with always-on adaptive thinking and a one-million-token context window, aimed at hundreds of millions of users. Every day it stays dark is revenue and mindshare ceded to rivals and to open-weights alternatives. So the company is caught between moving fast to restore and ship, and moving carefully to avoid a second recall. How it resolves that tension is, in effect, the roadmap.

This is where most "next Claude model" coverage falls apart, by blending fact and guesswork. Here is the clean separation.

Confirmed (from Anthropic or its filings):

• Opus 4.8 is the current general-purpose flagship, released 28 May 2026.
• Fable 5 and Mythos 5 launched 9 June 2026, were suspended 12 June 2026, and were restored 1 July 2026.
• Claude Sonnet 5 launched 30 June 2026, with near-Opus performance at a lower price.
• Project Glasswing has been expanded to 150 additional organisations across more than 15 countries.

Rumoured or speculative (treat with caution):

• Claude Opus 5. The active rumour as of late July 2026. A research checkpoint codenamed "Honeycomb EAP" surfaced briefly in Cursor around 9 July, reportedly showing a 1 million-token context window and an "xhigh" reasoning-effort tier with a safety fallback to Opus 4.8. A claude-opus-5 string was also reportedly spotted on Google Vertex AI's Model Garden on 14 July, unconfirmed by either Anthropic or Google. Multiple outlets pointed to 23 July as a likely launch date; it did not happen on schedule, and Anthropic has issued no official announcement or system card as of 24 July.
• A clean Claude 5 generation, distinct from Opus 5. Prediction markets have floated a 2026 window, but Anthropic has not announced it and has so far always preferred point releases and product-line names (Sonnet, Opus, Fable) over a whole-number generational jump.
• A Haiku refresh on the latest generation. Logical, given Haiku 4.5 shipped in late 2025, but unannounced.

The honest position is that, while Sonnet 5 and Fable 5's return are now settled fact, no Opus 5 release is confirmed as of 24 July 2026 despite the volume of leaks. Anyone quoting you a hard Opus 5 launch date is reading tea leaves, however credible the underlying leak looks. This piece will be updated the moment Anthropic actually announces it.

It helps to understand why the rumour cycle is so persistent and so wrong. Three recurring sources feed it. The first is accidental disclosure, like the Fennec identifier that surfaced in a Google Vertex AI log in early February 2026 and was immediately read as Sonnet 5; it shipped two weeks later as Sonnet 4.6. The lesson is that an internal checkpoint name tells you a model exists, not what it will be called or how big a step it represents. The second source is prediction markets, which put real but soft probabilities on windows like "Claude 5 by mid-2026." Those are useful as sentiment, not as schedule. The third is simple pattern-matching journalism, where "Opus reached 4.8, so 5 must be next" gets written up as if cadence were a promise.

Anthropic's own behaviour is the best corrective. The company has repeatedly had the option to badge a release as a new generation and declined, preferring the decimal. That is a deliberate signal: it reserves the whole number for a step it considers genuinely generational, and it is in no rush to spend that signal. So when you read "Claude 5 is coming," the right mental translation is "a capable new checkpoint is plausibly coming, and Anthropic will most likely call it 4-point-something."

Two of the candidates this section originally ranked have since shipped: Fable 5 and Mythos 5 were restored on 1 July, and Claude Sonnet 5 launched 30 June rather than lagging as a rumoured "Sonnet 4.8." Here is the ranking updated for what is actually still open.

1. Claude Opus 5. By far the most likely next new model, and the most heavily leaked. The Honeycomb EAP checkpoint, a reported Vertex AI Model Garden sighting, and the general density of coordinated leak activity around 23 July all point to a model that is finished or very close to it internally. This is our top pick for "what actually ships next."

2. A hardened Mythos-class point release (Fable 5.1 or similar). Now that Fable 5 and Mythos 5 are back online, a follow-up refinement addressing whatever safeguard changes made restoration possible is a reasonable medium-term candidate, though nothing has been announced.

3. A Haiku refresh. Logical, given Haiku 4.5 shipped in late 2025 and every other tier (Opus, Sonnet, the Mythos-class) has since moved, but there remains no signal on timing.

4. A clean "Claude 5" generational rebrand. Still the least likely near-term move. Anthropic has now shipped Sonnet 5 as a product-line version number, not a Claude-wide generational reset, which is exactly the pattern we predicted in June: version numbers climb within Opus, Sonnet, Haiku and Mythos independently, rather than the whole Claude brand jumping to "5" at once. Opus 5, if and when it lands, will most likely follow the same logic.

A word on Fable as a name, because it is still easy to misread. Fable is not a tier in the way Haiku, Sonnet and Opus are. It sits on the Mythos-class lineage above Opus, and the "5" in Fable 5 and Mythos 5 was a Mythos-class version number, not a Claude generation number. That is precisely why Fable 5 shipping did not imply a Claude 5, and why Sonnet reaching "5" separately, on its own product-line clock, is the more instructive precedent for how Opus 5 will likely be named and scoped.

What would Opus 5 actually need to improve on, if the leaks are accurate? A 1 million-token context window would bring Opus into parity with what Fable 5 already offered, an "xhigh" reasoning tier would sit above today's effort settings for the hardest agentic and research tasks, and a safety fallback to Opus 4.8 would be a direct, structural answer to the Fable 5 recall: build a way to step down to an already-cleared model into the architecture itself, rather than relying solely on pre-release review to catch every risky capability before launch.

For practical planning, the takeaway for builders is unchanged in shape even as the specifics moved: design for the track, not the number. If you depend on the absolute frontier, watch Opus closely right now and keep Opus 4.8 as your stable fallback until Opus 5 is confirmed and independently evaluated. If you depend on cost-efficiency, Sonnet 5 is now your answer rather than a rumour to wait on. If you depend on the very top of the frontier with fewer guardrails, the Mythos-class remains the most regulation-exposed track, as the June recall demonstrated.

Whatever ships next, money is not the bottleneck. Anthropic's capital position in 2026 is extraordinary.

In February 2026 the company raised around £23bn ($30bn) in a Series G round at a roughly £287bn ($380bn) post-money valuation. Just months later it raised about £49bn ($65bn) in a Series H round at a roughly £728bn ($965bn) post-money valuation, closing in on the £755bn ($1tn) mark. The Series H was co-led by a deep bench including Altimeter, Coatue, Dragoneer, Greenoaks, Sequoia, Capital Group and GIC, with strategic memory and infrastructure partners such as Samsung, SK Hynix and Micron also taking part.

On 1 June 2026, Anthropic confidentially filed a draft S-1 with the SEC, opening the door to an IPO that the market widely expects to target a £755bn ($1tn)-plus valuation.

Revenue has tracked the valuation. Run-rate revenue was reported at around £11bn ($14bn) in February 2026 and crossed roughly £35bn ($47bn) in May 2026, having grown more than tenfold annually in each of the past three years. That growth is the engine behind the accelerating release cadence, more revenue funds more compute funds faster models.

The two strategic backers are central. Amazon's roughly £6bn ($8bn) investment is now reported to be worth well over £53bn ($70bn), with its overall stake estimated in the mid-to-high teens of percent. Google holds roughly 14% in straight equity, contractually capped near 15%, worth on the order of £102bn ($135bn) at the current valuation. Reporting in 2026 even suggested that a large share of Google's and Amazon's headline AI profits traced back to their Anthropic stakes rather than their core businesses, a measure of how much investor interest now rides on the next Claude model performing.

Why does any of this matter for predicting the next model? Because the funding structure removes the usual excuse for delay. A lab that is capital-constrained ships when it can afford to. Anthropic is the opposite case: it has more committed capital and compute than it can deploy at once, and a pending IPO that rewards visible momentum. That combination pushes towards a steady drumbeat of releases, restored access and point-bumps, because each one is a proof point for the public-market story. The recall complicates the narrative, but it does not change the underlying incentive to keep shipping.

The IPO filing adds a subtler pressure. Once a company is in a registration process, every product decision is also a disclosure decision, and a recalled flagship is exactly the kind of risk that ends up described in an S-1. That gives Anthropic a strong reason to resolve the Fable 5 and Mythos 5 situation cleanly and publicly before any listing, which is another argument for restoration being a near-term priority over a flashy new generation. Investors want to see a frontier lab that can both reach the frontier and operate it within the rules.

It is also worth keeping the figures honest. The conversions here put pounds first with dollars in brackets at an indicative rate, and the run-rate revenue numbers are reported figures rather than audited results, so they should be read as direction and scale rather than precise accounting. What is not in doubt is the order of magnitude: a near-trillion-pound valuation, tens of billions in annualised revenue, and two hyperscaler backers whose own results now move with Anthropic's. That is the financial backdrop against which the next Claude model will land.

The other precondition for a new frontier model is raw compute, and Anthropic has locked in a multi-cloud, multi-gigawatt position that is among the largest in the industry.

In late 2025 Google committed up to roughly £30bn ($40bn) in cash and TPU compute, giving Anthropic access to up to one million TPUs and creating multi-gigawatt capacity through 2026, a deal later framed alongside a 3.5GW Broadcom-built TPU buildout.

Then on 20 April 2026, Amazon announced it would invest up to around £19bn ($25bn) more in Anthropic, with Anthropic committing over £75bn ($100bn) across ten years to AWS and the two deploying up to 5GW of new Trainium capacity. Roughly 1GW of combined Trainium2 and Trainium3 capacity is expected online by the end of 2026. Industry observers called it the largest single-customer endorsement of non-Nvidia silicon in AI training history.

The takeaway for the roadmap is simple. Anthropic has the silicon to train and serve the next several generations. If the next Claude model slips, it will be because of safety, regulation or product strategy, not because the company ran out of chips.

The multi-cloud, multi-silicon shape of these deals matters in its own right. By committing to both AWS Trainium and Google TPUs at gigawatt scale, Anthropic reduces its dependence on any single supplier, including Nvidia, and gives itself parallel training and serving capacity. For release cadence, parallel capacity is the quiet enabler: it lets the company train candidate checkpoints, run extensive evaluations and serve production traffic at the same time, which is exactly the workflow behind a six-to-eight-week point-release rhythm. A serial, single-cluster lab simply cannot move that quickly.

There is a strategic wrinkle worth naming. Massive serving capacity is what made it commercially sensible to offer a frontier model with always-on adaptive thinking and a one-million-token context window to hundreds of millions of users, as Fable 5 was. The same capacity is what makes a fast restoration feasible once the regulatory issue is resolved, because the infrastructure to serve it never went away. In other words, the hard, slow part, building the compute, is already done. The constraint on the next model is now almost entirely about permission and safety rather than physics.

Safety policy is now a first-class input to the release schedule, not an afterthought. Anthropic published Responsible Scaling Policy version 3.0, effective 24 February 2026, a comprehensive rewrite that introduces Frontier Safety Roadmaps with detailed safety goals and Risk Reports that quantify risk across deployed models.

The RSP is built on proportional protection, with AI Safety Level (ASL) standards that tighten as capability rises. ASL-3 safeguards, triggered when a model could meaningfully help someone with basic technical skill create or deploy CBRN weapons, involve enhanced security and deployment controls including stronger protection of model weights. ASL-4 measures are not yet fully written; Anthropic's commitment is to define them before reaching ASL-3, and they may require assurance methods that are still open research problems, such as using interpretability to show mechanistically that a model will not behave catastrophically.

Project Glasswing is the cyber arm of this trajectory. It pairs the Mythos-class with critical-infrastructure defenders to find and patch vulnerabilities before attackers can. Launch partners included AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, and partners have reportedly surfaced more than 10,000 high or critical-severity flaws, including some in every major operating system and browser.

The recall is the friction point where this safety machinery met government power. Anthropic's whole Fable design was meant to keep offensive cyber behind classifiers that fall back to Opus 4.8. The directive treated a reported gap as reason to pull the product entirely. For the next model, expect the safety and classifier story to be even more central, because regulators are now watching the cyber capability of every Mythos-class release closely.

The RSP v3.0 machinery is itself a roadmap signal, because it dictates what Anthropic must do before it can ship at higher capability. The commitment to write ASL-4 measures before reaching ASL-3 is, in effect, a self-imposed gate that can slow a generational leap. If the next model would cross into territory that triggers higher ASL standards, then by Anthropic's own policy it cannot ship until the corresponding safeguards and, for ASL-4, the underlying assurance research exist. That is one structural reason to expect point releases within the current capability band rather than a sudden jump: staying inside a known ASL tier is faster than clearing a new one.

Glasswing also reframes how to read the cyber capability that caused the recall. The same vulnerability-finding ability that worries export-control officials is the ability Anthropic is using, through partners, to patch flaws in critical infrastructure at scale, reportedly more than 10,000 high or critical issues including some in every major operating system and browser. That is the dual-use bind in a single programme: the capability that makes a model dangerous to release widely is the capability that makes it valuable to defenders. The next Mythos-class release will have to thread that needle more visibly than Fable 5 did, probably by leaning harder on the safeguarded-public versus vetted-partner split that the recall stress-tested.

The publication of Frontier Safety Roadmaps and quantified Risk Reports is the other thing to watch. These are the documents in which Anthropic effectively pre-announces the kinds of capability it is preparing to handle and the safeguards it is building for them. They are not a product roadmap, but they are the closest thing to an official signal about the direction of the next frontier model, and they are worth reading alongside any release rumour as a sanity check on what is actually plausible.

Anthropic does not set its cadence in a vacuum. OpenAI is on a parallel, equally fast track, and the two roadmaps push each other.

OpenAI released GPT-5.5, codenamed Spud, on 23 April 2026, with API access the following day. GPT-5.6 has since launched and now powers OpenAI's ChatGPT Work agent platform, announced 9 July 2026. GPT-6 has been signalled by OpenAI's leadership as in training and expected to arrive faster than the gap between GPT-4 and GPT-5, with prediction markets pricing a Q3-to-Q4 2026 base case rather than imminent.

So both labs are in the same posture: rapid point releases at the top of the range, with a true generational model (Claude 5, GPT-6) plausibly later in 2026 but unconfirmed. The competitive read is that neither can afford a long gap, which supports the case for a near-term Anthropic release, whether that is a restored Fable, a hardened Mythos-class variant or an Opus bump. We track the parallel question for OpenAI in our sister piece, What's the Next GPT Model?.

There is an interesting asymmetry in how the two companies talk about generations. OpenAI's leadership has openly signalled that GPT-6 is in training and will arrive faster than the GPT-4-to-GPT-5 gap, effectively pre-selling the next whole number. Anthropic does the opposite, downplaying the generational label and shipping decimals. That is partly branding and partly philosophy: Anthropic's public identity is built around safety and measured deployment, so a quiet 4.8 fits the story better than a triumphant 5. For readers trying to predict the next model, the practical effect is that OpenAI's roadmap is louder and Anthropic's is quieter, which makes Anthropic's harder to forecast even when the underlying cadence is similar.

The recall sharpens the comparison in one specific way. Anthropic explicitly argued that the cyber capability behind the Fable 5 suspension is already available in GPT-5.5 and used daily by defenders. If regulators take that argument seriously, the next move could be industry-wide scrutiny rather than a single-company recall, which would slow everyone, OpenAI included. If they do not, Anthropic is left carrying a constraint its main rival does not, which would be a real competitive disadvantage for its frontier track. Either way, the next Claude model is now partly a regulatory question, and that is the clearest difference between the two roadmaps in mid-2026.

Our June prediction called the restoration of Fable 5 and Mythos 5 as the next visible move, and that is exactly what happened on 1 July. Here is our updated, clearly-flagged-as-speculation call for what comes next.

We expect Claude Opus 5 to be the next genuinely new model, and we think it is close. The Honeycomb EAP leak, the reported Vertex AI Model Garden sighting, and the sheer density of coordinated leak activity around 23 July all point to a model that is finished or nearly finished internally, with the remaining delay more likely about final safety sign-off and go-to-market timing than fundamental readiness. A launch within the next few weeks of 24 July looks more likely than not, though "Opus ships on Thursdays" has already failed to predict the exact date once.

On specification, if the leaked details hold up, expect a 1 million-token context window, a new "xhigh" extra-high reasoning-effort tier sitting above today's options, and a safety fallback design that can route sensitive requests back down to the already-safety-tested Opus 4.8 rather than shipping every capability at full autonomy from day one. That fallback pattern would also be a sensible lesson learned directly from the Fable 5 recall: build the safety off-ramp into the architecture before regulators force one.

We do not expect Anthropic to badge this release as a clean "Claude 5" generation. The naming pattern has held throughout 2026 even under commercial pressure, and Opus 5 as a product-line name (Opus, not Claude, hitting version 5) is consistent with how Sonnet 5 and the Mythos-class Fable/Mythos 5 names were handled: version numbers climb within a product line, not across the whole Claude brand.

The next Claude model is being shaped less by Anthropic's labs than by a government directive. The cadence said a new flagship was due imminently. The recall said the next priority is getting the last one back online and proving it is safe.

What is confirmed is a fast-moving, well-funded, compute-rich lab with a clear safety framework and a suspended frontier model it wants to restore. What is rumoured is everything with a 5 in the name. The new roadmap is real, but it runs through regulation now, and that is the genuinely new thing about Anthropic's 2026.