Anthropic shipped Sonnet 4.5 on Sept. 29, 20251. The announcement opened with a claim, stated flat. "Claude Sonnet 4.5 is the best coding model in the world. It's the strongest model for building complex agents." Its SWE-bench Verified score came in at 77.2%, averaged over 10 trials on the full 500-problem set. That release sat in the middle of a run, from Claude 3.5 Sonnet in mid-2024 to Sonnet 5 in June 20262. Nine named models, across two years, most packed into thirteen months. Behind the names sat a smaller set of choices. Who defines Claude's direction, and who gets paid to keep the pace funded. Jared Kaplan, Chris Olah and Jan Leike answer the first. Google, Amazon, Microsoft, Nvidia, AMD and Nscale answer the second3.
A Cadence That Kept Accelerating
Claude 3.5 Sonnet launched June 20, 2024, then got an upgrade beside Claude 3.5 Haiku that October2. A hybrid model, Claude 3.7 Sonnet, followed in February 2025, with a thinking mode users could switch on. May 22, 2025 brought Claude 4, split at once into Opus 4 and Sonnet 4. Opus 4.1 followed that August. Sonnet 4.5 landed in September, Haiku 4.5 on Oct. 15, Opus 4.5 on Nov. 241. Three major releases inside ten weeks closed the year.
2026 moved faster. Opus 4.6 shipped Feb. 5, Sonnet 4.6 twelve days later, Opus 4.7 on April 16, Opus 4.8 on May 282. Gaps that once ran a full quarter shrank to weeks. Sonnet 5 closed the sequence June 30, 2026. Set that pace against Claude's early public image, the careful option next to faster rivals. A schedule this dense trades some of that image for speed. The bet is that the safety work keeps up inside the building even when the version numbers say hurry.
Anthropic packed three Claude releases into 10 weeks to close 2025, then shipped faster in 2026
Claude model releases by date, from Claude 4 on May 22, 2025 to Sonnet 5 on June 30, 2026
Each pin is one Claude release on a rail spaced by real time, so pins that crowd together mark releases that came weeks apart.
Sources: Anthropic [1]; Wikipedia [2]. Chart by Kelly Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| May 22, 2025 | Opus 4 and Sonnet 4. Claude 4: Opus 4 and Sonnet 4 |
| Aug. 2025 | Opus 4.1 |
| Sept. 29, 2025 | Sonnet 4.5. Scored 77.2% on SWE-bench Verified |
| Oct. 15, 2025 | Haiku 4.5 |
| Nov. 24, 2025 | Opus 4.5. The third release inside ten weeks |
| Feb. 2026 | Sonnet 4.6. Twelve days after Opus 4.6 |
| Feb. 5, 2026 | Opus 4.6 |
| April 16, 2026 | Opus 4.7 |
| May 28, 2026 | Opus 4.8 |
| June 30, 2026 | Sonnet 5. Closes the sequence |
Customers saw the coding gains first. Michael Truell, who runs Cursor, reported the result from his own product. "We're seeing state-of-the-art coding performance from Claude Sonnet 4.5, with significant improvements on longer horizon tasks."1 Scott Wu, cofounder and chief executive of the company behind Devin, gave numbers. "For Devin, Claude Sonnet 4.5 increased planning performance by 18% and end-to-end eval scores by 12%—the biggest jump we've seen since the release of Claude Sonnet 3.6." Read what both men measure. Longer tasks and planning, rather than raw answer quality.
Kaplan, Olah and the Research Leadership
Jared Kaplan, an Anthropic co-founder, is chief science officer, which puts training strategy and scaling-law work under him3. He co-wrote the original scaling-laws papers that taught the field to predict what a model can do from the compute it gets. His seat ties the release pace to research he helped start before Anthropic existed.
Chris Olah, another co-founder, leads interpretability, the work of explaining what happens inside a model's weights rather than grading its answers3. Amanda Askell works on Claude's character, which sits apart from safety work narrowly defined. Her focus is the personality and values a shipped model shows in conversation. Few frontier labs make that a formal job at all.
Jan Leike joined to co-lead Alignment Science after leaving OpenAI's superalignment effort35. His resignation letter in 2024 said why. "Over the past years, safety culture and processes have taken a backseat to shiny products," Leike wrote, adding that he had "gradually lost trust" in OpenAI's leadership. Landing at Anthropic reads as a researcher picking the lab whose stated priorities matched his own. Hiring him reads as a company treating that choice as a credential worth paying for.
Six Partners, One Supply Chain
Anthropic's compute list widened sharply across 2025 and 2026, away from the single-cloud habit that defines some rivals. Google's TPU deal gives it access to up to one million Tensor Processing Units, dated October 20253. Amazon stays the main cloud provider through its Trainium line, backed by roughly $8 billion put in between September 2023 and November 2024. Microsoft and Nvidia struck a joint Azure capacity purchase worth $30 billion that November.
2026 added three more names. Akamai signed a $1.8 billion cloud deal in May, the same month Anthropic got data-center access through xAI's Colossus site3. AMD followed in July with 2 gigawatts of MI450 GPUs, plus an investment in Anthropic worth up to $5 billion. Nscale closed the list in August 2026, roughly $45 billion tied to about 460 megawatts in West Virginia. Six partners inside two years describes a company that refuses to lean on one chip design or one cloud. Call it a hedge against price leverage and supply shocks, both of which a company scaling this fast can ill afford.
The Revenue the Cadence Bought
Every release and every compute deal traces to one number. Revenue reached a $65 billion annual pace by late July 2026, up from $9 billion at the end of 2025, per TechCrunch reporting matched the same day elsewhere4. Claude Code sits inside that growth as a named driver of enterprise use. The company held back a Claude Code revenue figure as of this writing.
Anthropic's revenue run rate rose from $9 billion at the end of 2025 to $65 billion by late July 2026
Anthropic's annualized revenue pace, in billions of dollars, per TechCrunch
The blue column is the run rate at the end of 2025, and the magenta column is the run rate by late July 2026.
Source: TechCrunch [4]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| End of 2025 | $9 billion | |
| Late July 2026 | $65 billion | Claude Code is a named driver of enterprise use |
Kaplan's scaling research, Olah's interpretability work and Leike's alignment focus are the argument Anthropic makes to buyers wary of frontier risk. The compute list is the proof that argument scales past any one vendor slowing supply. Research credibility and hardware diversity hold up a $65 billion run rate together. Model quality alone would hold up less.
What Nine Releases Signal
A pace moving from quarterly to near-monthly in eighteen months forces a choice. Either safety review speeds up to match the schedule, or the schedule slows to match safety review. Anthropic takes the first answer in public, crediting internal alignment work strong enough to absorb the pace. Whether that holds matters more than the version numbers. Sonnet 5 and whatever follows will meet buyers who want evidence rather than assurance.
Chris Painter, director of policy at METR, the nonprofit that tests frontier models for catastrophic risk, reviewed an early draft of Anthropic's revised scaling policy before it published in February 2026. He called the revision understandable, then named the problem it signals. The change shows Anthropic "believes it needs to shift into triage mode with its safety plans, because methods to assess and mitigate risk are not keeping up with the pace of capabilities," Painter told TIME6. Buyers reading nine releases as proof that review keeps pace are hearing Anthropic's own account of itself. An outside evaluator reads the same policy change as evidence that review lags the schedule it exists to police.
The compute spread tells a parallel story, about risk moved from model behavior to hardware. Six partners, across three chip designs and two continents of capacity, cut the odds that one disruption stalls the pace Kaplan's team set. Anthropic built backup into its hardware supply about as deliberately as it built backup into its safety research. One habit, done twice.
By the numbers
- A SWE-bench Verified score of 77.2% arrived with Claude Sonnet 4.5, averaged across 10 trials1.
- Revenue run rate reached $65 billion by late July 2026, up from $9 billion at the end of 20254.
- Nine named Claude generations run from 3.5 Sonnet through Sonnet 5, June 2024 to June 30, 20262.
- Microsoft and Nvidia committed $30 billion in joint Azure capacity in November 20253.
- One million custom Tensor Processing Units sit available under the Google partnership, dated October 20253.
- Nscale's compute agreement runs roughly $45 billion, tied to about 460 megawatts in West Virginia3.
- Two gigawatts of MI450 GPUs came from AMD in July 2026, with an AMD investment of up to $5 billion3.
What to watch
Sonnet 5's reception among enterprises, set against Claude Code's share of revenue, will show whether the faster pace keeps winning paying accounts or starts to strain trust4. A published Claude Code figure would let outside analysts test the growth against one product rather than the total. More compute partners, or a pause, would show whether the spread hit its intended size or keeps widening to match a pace still gaining speed3.
Sources
- Anthropic, "Claude Sonnet 4.5," Anthropic, Sept. 29, 2025, https://www.anthropic.com/news/claude-sonnet-4-5
- Wikipedia contributors, "Claude (language model)," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Claude_(language_model)
- Wikipedia contributors, "Anthropic," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Anthropic
- TechCrunch Staff, "Anthropic's annualized revenue surges to $65B," TechCrunch, Aug. 17, 2026, https://techcrunch.com/2026/08/17/anthropics-annualized-revenue-surges-to-65b/
- Wikipedia contributors, "Jan Leike," Wikipedia, accessed Sept. 4, 2026, https://en.wikipedia.org/wiki/Jan_Leike
- Billy Perrigo, "Exclusive: Anthropic Drops Flagship Safety Pledge," TIME, Feb. 24, 2026, https://time.com/7380854/exclusive-anthropic-drops-flagship-safety-pledge/