What is the biggest software developer news from July 20 to 26, 2026?

The week belonged to model economics and agent tooling. On July 24, Anthropic released Claude Opus 5, priced the same as Opus 4.8 at $5 per million input and $25 per million output tokens but close to frontier Fable 5 intelligence and a new state of the art on coding benchmarks like Frontier-Bench and GDPval-AA. On July 21, Google shipped three cheaper Gemini Flash models, including a restricted Flash Cyber security model, while its long-promised Gemini 3.5 Pro flagship stayed in partner testing. On July 22, Cursor launched Cursor Router, which picks the best model per request and cut costs 30 to 60 percent in its tests. On July 23, OpenAI brought full-duplex GPT-Live voice into Codex so developers can drive multiple coding agents hands-free. On the security side, a public exploit for SharePoint CVE-2026-50522 was used within hours to steal machine keys, and a poisoned PyPI release of mrmustard stole cloud and SSH credentials at import time.

What is Claude Opus 5 and how is it different from Opus 4.8?

Claude Opus 5 is Anthropic's new Opus-class model, released July 24, 2026. Anthropic says it comes close to the frontier intelligence of Claude Fable 5 at half the price, and it is priced the same as its predecessor Opus 4.8 at $5 per million input tokens and $25 per million output tokens. It sets a new state of the art on coding and knowledge-work evaluations like Frontier-Bench and GDPval-AA, scoring 43.3 percent on Frontier-Bench v0.1 versus 18.7 percent for Opus 4.8 and 33.7 percent for Fable 5, though it stays behind Mythos 5 on cybersecurity tasks. It becomes the default model on Claude Max and the strongest model on Claude Pro. Also shipping: a Fast mode running about 2.5 times faster at twice the price, automatic fallback routing on the API, and mid-conversation tool changes that no longer invalidate the prompt cache. It carries a 1 million token context window and no data retention requirements. Developers can use it as claude-opus-5 on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

What did Google announce with Gemini on July 21, 2026?

On July 21, 2026, Google released three new Gemini models: Gemini 3.6 Flash, a cheaper general-purpose workhorse that uses roughly 17 percent fewer tokens per task; Gemini 3.5 Flash-Lite, an even cheaper low-latency model; and Gemini 3.5 Flash Cyber, a security-tuned model for vulnerability analysis that is restricted to governments and vetted partners through a limited pilot. Google also said it has begun its most ambitious pre-training run yet for Gemini 4. Notably absent was Gemini 3.5 Pro, the flagship Google previewed at I/O and promised for June, which remains in partner testing. Independent benchmarking found the new Flash models cost less per task than comparable OpenAI and Anthropic offerings, and the restricted Flash Cyber release mirrors the gated approach Anthropic took with its Claude Mythos security model.

What is Cursor Router?

Cursor Router, launched July 22, 2026, is a request-level classifier that inspects each coding query and dispatches it to the model best suited for that task, rather than sending everything to one default model. Cursor says it trained the router on more than 600,000 live requests and evaluated it across millions, and that it delivers frontier-quality output at up to 60 percent lower cost in A/B tests, with 30 to 50 percent savings reported by early-access enterprise accounts. It offers three modes, Intelligence, Balance, and Cost, that move you along the cost-intelligence tradeoff. The model pool includes Fable 5, Opus 4.8, GPT-5.6 Sol, Grok 4.5, and Cursor's own Composer. Router is available for Teams and Enterprise plans across desktop, web, iOS, CLI, and the SDK, on by default for Teams, and admin-enabled for Enterprise. Balance and Intelligence modes bill at the routed model's rate, and Grok 4.5 cannot be excluded from the pool.

What is the SharePoint CVE-2026-50522 exploit and why does patching alone not fix it?

CVE-2026-50522 is a critical Microsoft SharePoint Server deserialization vulnerability that allows remote code execution. Microsoft patched it in the July 2026 Patch Tuesday but did not initially mark it as exploited. On July 20, 2026, offensive security firm watchTowr identified public proof-of-concept exploit code, and within hours its honeypot network captured attacks that successfully compromised on-premises SharePoint servers. CISA added the flaw to its Known Exploited Vulnerabilities catalog on July 22 with a July 25 remediation deadline. The dangerous part is that attackers use the flaw to steal SharePoint machine keys in a single request, which lets them forge authentication tokens and keep access even after the server is patched. Security teams are advised to patch, then rotate machine keys and any other exposed credentials, because applying the update alone does not evict an attacker who already grabbed the keys.

Dev Weekly Jul 20-26, 2026: Anthropic Ships Claude Opus 5 at Half the Price, Google Floods the Market With Three Gemini Flash Models, Cursor Router, OpenAI Voice-Drives Codex, SharePoint Flaw Exploited

The theme this week was cheaper intelligence and faster agents. Anthropic shipped Claude Opus 5, a model it says lands near frontier Fable 5 quality at half the price and the same rate as the model it replaces, and made it the new default on its consumer tiers. Google answered its stalled flagship with three Gemini Flash models tuned for cost, including a locked-down cybersecurity variant, and quietly said Gemini 4 pre-training has started. Cursor stopped asking developers to pick a model at all and shipped a router that decides per request, claiming frontier output for 30 to 60 percent less. OpenAI put full-duplex voice into Codex so you can run several coding agents while your hands stay off the keyboard.

Security did not let anyone celebrate for long. A public exploit for a SharePoint deserialization bug went from proof of concept to real-world machine-key theft in hours, and patching alone does not clean it up. A hijacked PyPI package poisoned a quantum library with an import-time credential stealer that reaches for SSH keys, cloud secrets, and Kubernetes configs. Anthropic answered the moment by shipping a Claude Security plugin that puts a team of agents to work hunting bugs in your own codebase. Add Vercel’s first scheduled Next.js security release, a fresh wave of AI-era layoffs, and billions more into physical AI and inference chips, and here is everything that mattered.

Top Stories This Week

Anthropic Ships Claude Opus 5 at Half the Price of Fable 5 -

On July 24, Anthropic released Claude Opus 5, a model it says comes close to the frontier intelligence of Claude Fable 5 at half the price. The headline for developers is the pricing: Opus 5 costs the same as its predecessor Opus 4.8, $5 per million input tokens and $25 per million output tokens, while setting a new state of the art on coding and knowledge-work benchmarks. On Frontier-Bench v0.1, an agentic terminal coding test, Anthropic reports Opus 5 scores 43.3 percent, more than double Opus 4.8’s 18.7 percent and ahead of Fable 5’s 33.7 percent, at a lower cost per task. It stays behind the gated Mythos 5 on cybersecurity work. Opus 5 is now the default model on Claude Max and the strongest model on Claude Pro.

As VentureBeat reported, the launch signals how the AI race is shifting from raw capability to the economics of daily use. Also shipping the same day: a Fast mode that runs roughly 2.5 times faster at twice the base price, automatic fallback routing on the API for content its safety classifiers flag, and mid-conversation tool changes that no longer invalidate the prompt cache, a small feature agent developers will appreciate more than any benchmark. Opus 5 carries a 1 million token context window, no data retention requirements for general access, and is available as claude-opus-5 on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, plus inside tools like Kiro and Lovable.

Google Floods the Market With Three Gemini Flash Models as Its Flagship Slips -

On July 21, Google shipped three new Gemini models at once: Gemini 3.6 Flash, a cheaper general-purpose workhorse that Google says uses about 17 percent fewer tokens per task; Gemini 3.5 Flash-Lite, an even cheaper low-latency option for agent loops; and Gemini 3.5 Flash Cyber, a security-tuned model for vulnerability analysis that goes only to governments and vetted partners through a limited pilot. In the same update, Google said it has begun what it calls its most ambitious pre-training run yet for Gemini 4. All three Flash models are live now in Google AI Studio and the Gemini API.

What Google did not ship is the model everyone is waiting for. As The State of AI noted, Gemini 3.5 Pro, the flagship previewed at I/O and once promised for June, is still stuck in partner testing, and independent benchmarking found the new Flash line cheaper per task than comparable OpenAI and Anthropic offerings. The AI Herald breakdown makes the strategic read clear: Google is willing to fragment its lineup to win specific use cases, and the restricted Flash Cyber release copies the gated playbook Anthropic used for its Claude Mythos security model. For developers, the practical takeaway is that the cheap tiers are getting genuinely good, but do not architect around a Pro model that still has no public date.

Cursor Launches a Router That Picks the Model for You -

On July 22, Cursor launched Cursor Router, a request-level classifier that reads each coding query and sends it to the model best suited for that task instead of routing everything to a single default. Cursor says it trained the router on more than 600,000 live requests, evaluated it across millions in an online A/B test, and that it delivers frontier-quality output at up to 60 percent lower cost. It ships with three modes, Intelligence, Balance, and Cost, that let teams pick where they sit on the cost-intelligence tradeoff, and the model pool spans Fable 5, Opus 4.8, GPT-5.6 Sol, Grok 4.5, and Cursor’s own Composer.

As MarkTechPost detailed, early-access enterprise accounts saw 30 to 50 percent savings when Cursor compared their actual spend against the same traffic priced entirely at Opus 4.8 rates. Two caveats matter for buyers. Balance and Intelligence modes bill at the routed model’s rate, so your unit cost varies with each routing decision rather than settling at a flat price, and Grok 4.5 is required as a price-efficient option, so it cannot be removed from the pool via the block list. Router is available across desktop, web, iOS, CLI, and the SDK, on by default for Teams and admin-enabled for Enterprise. Individual Pro plans keep manual model selection for now.

OpenAI Voice-Drives Codex With Full-Duplex GPT-Live -

On July 23, OpenAI brought its full-duplex GPT-Live voice model into the ChatGPT desktop app on macOS and Windows, wiring it directly into Codex and ChatGPT Work. Because GPT-Live listens and speaks at the same time, developers can queue several coding agents, check on their progress, review pull requests, and redirect active jobs by talking, without stopping to type or switch windows. GPT-Live manages the conversation and echoes small acknowledgments while background models handle the actual planning and code generation, and on macOS a screen context feature lets you say “take a look at this” to share the frontmost window.

As 9to5Mac reported, the feature rolled out globally the same day as desktop build 26.715 and is limited to paid Plus, Pro, Business, Edu, and Enterprise plans, with remote task checks available from iOS. One honest limitation: there is still no developer API for GPT-Live, so you cannot build custom voice agents on this exact stack; the older Realtime API remains the buildable path. It landed the same week Anthropic shipped its own voice mode, a sign both labs are betting that once agents run unattended, the real interface problem becomes supervising several jobs at once.

SharePoint CVE-2026-50522 Goes From Public PoC to Machine-Key Theft in Hours -

On July 20, attackers began exploiting a critical SharePoint flaw, CVE-2026-50522, to steal machine keys and keep access even after servers are patched. Microsoft fixed the deserialization remote code execution bug in the July Patch Tuesday but did not initially mark it as exploited. That changed when offensive security firm watchTowr identified public proof-of-concept exploit code on July 20 and, within hours, saw its global honeypot network capture attacks that successfully compromised on-premises SharePoint deployments. The published exploit delivers a malicious .NET payload through a forged security token posted to SharePoint’s sign-in endpoint.

As Security Affairs reported, CISA added CVE-2026-50522 to its Known Exploited Vulnerabilities catalog on July 22 with a July 25 remediation deadline. The catch that makes this worse than a normal RCE: attackers pull the SharePoint machine keys in a single request, and those keys let them forge authentication tokens for persistent access long after the update lands. Security teams are warned that patching is not enough. You have to rotate machine keys and any other credentials that may have been exposed, or an attacker who already grabbed the keys stays in. On-premises SharePoint remains one of the most reliably attacked pieces of enterprise software in 2026.

Developer Tools & Platforms

Anthropic Ships a Claude Security Plugin for Claude Code -

On July 22, Anthropic released the Claude Security plugin for Claude Code in beta, a multi-agent vulnerability scanner that runs entirely inside your existing session with no separate daemon. Install it with /plugin install claude-security@claude-plugins-official, run /claude-security, and pick one of three jobs: scan the whole repository, scan just a branch diff, pull request, or single commit, or turn confirmed findings into patch files you review and apply yourself. According to Anthropic’s documentation, a team of agents maps your architecture, builds a threat model, fans research agents across the code, and clears every finding through a three-voter adversarial panel before it reaches the report. Nothing is patched automatically, scans count against your plan’s token limits, and it needs Claude Code v2.1.154 or later with Python 3.9.6 on your PATH. It is the in-session version of Anthropic’s hosted Claude Security product, and it lands the same week a bank open-sourced a similar agentic bug hunter.

Vercel Patches Nine Next.js Vulnerabilities in Its First Scheduled Release -

On July 21, Vercel published the first release under its new monthly Next.js security program, patching nine vulnerabilities across four high and five medium severity issues in versions 16.2.11 (Active LTS) and 15.5.21 (Maintenance LTS). As the GitHub release notes spell out, the high-severity fixes include two server-side request forgery bugs, a denial of service in the App Router using Server Actions, and a middleware and proxy bypass in Turbopack apps using a single locale. Announced ahead of time as promised under the preannounced release model, the update slipped one day from its July 20 target. There are no backports for the end-of-life 13.x and 14.x lines, so teams stuck on those majors have to upgrade rather than patch. Run npm install next@16.2.11 or next@15.5.21 and redeploy.

Deno 2.9.4 Lands With a V8 Upgrade and Better Node Compatibility -

On July 23, Deno 2.9.4 shipped, a point release focused on fixes and performance for the JavaScript and TypeScript runtime. It upgrades V8 to 150.2.0 for the latest engine optimizations and security fixes, and continues closing the gap with Node.js: the ext/node layer adds a byteLength and length parameter to Buffer.indexOf, lastIndexOf, and includes, raw ChaCha20 support in crypto.createCipheriv, and proper argument validation for the node:sqlite backup() and deserialize() methods. Nothing headline-grabbing, but the steady Node-compat work is what keeps Deno viable as a drop-in for existing tooling. Run deno upgrade to update.

Security

Poisoned PyPI Release of mrmustard Steals Cloud and SSH Credentials at Import Time -

On July 24, a hijacked maintainer account published mrmustard 0.7.4 to PyPI with a credential stealer, a poisoned version of XanaduAI’s open-source quantum photonics library. The attacker took over a maintainer’s GitHub account, used the project’s own CI to steal its PyPI publishing token, then uploaded the malicious artifact straight to PyPI, so anyone reading the GitHub source saw nothing wrong. As the project’s security issue documents, the 0.7.4 tarball is identical to 0.7.3 except for a 258-line block injected into mrmustard/__init__.py, disguised as a _check_tf_compatibility() function. On every import mrmustard, it exfiltrates SSH private keys, AWS credentials, and Kubernetes configs to an attacker-controlled server, then installs three persistence mechanisms: a cron job, a shell RC hook, and a .pth file in site-packages so the payload runs on every Python invocation. Only 0.7.4 is malicious. If you imported it, treat every reachable credential as compromised, rotate from a clean host, and remove the persistence artifacts by hand. It is the same import-time, CI-token pattern that hit the @asyncapi npm packages a week earlier.

Etched Raises $300M at a $10.3B Valuation for Inference Chips -

On July 23, Etched said it raised $300 million at a $10.3 billion valuation, more than double its December mark and less than a month after emerging from stealth. The Series C was led by Sequoia with participation from a16z, Jane Street, Diffusion, and memory giant SK Hynix, and the company’s announcement calls it the highest valuation ever for a Sequoia-led Series C. Etched builds rack-scale inference clusters through deep co-design across the stack, and the cash goes toward scaling production and customer deployments at a new 80,000 square-foot Milpitas facility. It is another data point that the money around AI is flowing hard into the inference layer, not just the labs training frontier models.

Funding & Industry Deals

Travis Kalanick’s Atoms Raises $1.7B Led by a16z, With Uber Investing -

On July 22, Travis Kalanick’s robotics company Atoms raised $1.7 billion in a round led by Andreessen Horowitz, with Ben Horowitz joining the board and Bain Capital, Fifth Wall, and others participating. The most notable name on the cap table is Uber, the company Kalanick founded and was pushed out of in 2017. As The Next Web reported, Atoms is an industrial-AI and robotics company built in near-total stealth for years, and the equity comes paired with large credit facilities from JPMorgan, Goldman Sachs, Bank of America, Wells Fargo, and Barclays, the kind of firepower you raise to build heavy machines rather than apps. The company did not disclose a valuation, has shown little in the way of deployed robots at scale, and is betting it can digitize the physical industrial economy the way Uber digitized rides.

Meshy AI Raises $400M for 3D Generative Models -

This week, Meshy AI closed $400 million in Series B funding at a $1.5 billion valuation, led by Monolith Capital, IDG Capital, and Matrix Partners China. The Silicon Valley startup builds foundation models for AI-powered 3D generation, turning text and images into usable 3D assets for games, product design, and simulation. It was the second-largest round in a week that Crunchbase called varied and heavy, behind only Atoms, and it underscores that generative AI investment is spreading well past text and code into the 3D and physical-world pipelines that games and robotics teams depend on.

Layoffs: Monday.com, Patreon, Uber, and Magic Leap

Monday.com: On July 22, Monday.com said it is cutting about 630 employees, or 20 percent of its workforce, to restructure around its AI Work Platform. The Israeli software maker framed it as a shift from managing work to having humans and AI agents work together, not a cost cut, and said it expects $45 million to $55 million in charges. Layoffs.fyi data shows more than 122,000 tech roles cut so far in 2026, with a record 78 percent of companies citing an AI refocus.
Patreon: On July 23, Patreon laid off 20 percent of its staff, or 93 people, its largest cut since 2022. CEO Jack Conte said the core business is strong but the company had to adjust its cost structure and flatten its org chart, and stressed the cuts were not about replacing people with AI. Affected staff get at least 16 weeks of severance.
Uber: On July 22, Uber cut 10 percent of its customer service operations as it leans harder on AI, the first time Uber has explicitly tied layoffs to AI-driven efficiency. It asked remaining remote staff in the division to relocate to hub offices, and it follows a People-division cut two months earlier.
Magic Leap: On July 20, Magic Leap confirmed it is laying off 193 employees at its Florida headquarters per a WARN Act notice, as the AR company pivots to become a waveguide supplier and focuses on AI display glasses. The cuts span software, hardware, design, and senior leadership, effective October 1.

The Numbers That Matter

$5 / $25 Per million input/output tokens for Claude Opus 5, unchanged from Opus 4.8 but near Fable 5 quality
43.3% Opus 5’s Frontier-Bench v0.1 coding score, versus 18.7% for Opus 4.8 and 33.7% for Fable 5
30 to 60% Cost savings Cursor reports for its new per-request model Router
9 Vulnerabilities patched in Next.js 16.2.11 and 15.5.21, 4 high and 5 medium
$1.7 Billion Raised by Travis Kalanick’s Atoms, led by a16z with Uber participating
$10.3 Billion Etched’s new valuation after a $300 million Series C led by Sequoia
~630 Monday.com employees cut, about 20 percent of its workforce

Quick Hits

Magic Leap Layoffs - July 20. WARN notice confirms 193 job cuts as the AR firm pivots to being a waveguide supplier.
SharePoint CVE-2026-50522 Exploited - July 20. Public PoC used within hours to steal machine keys for persistent access.
Next.js Security Release - July 21. Vercel’s first scheduled release patches 9 vulnerabilities in 16.2.11 and 15.5.21.
Google Gemini Flash Trio - July 21. Gemini 3.6 Flash, 3.5 Flash-Lite, and a restricted Flash Cyber ship; Gemini 4 pre-training begins.
Cursor Router - July 22. Request-level model router for Teams and Enterprise, up to 60 percent cheaper in Cursor’s tests.
Claude Security Plugin - July 22. Anthropic’s multi-agent vulnerability scanner for Claude Code enters beta.
CISA Adds SharePoint Flaw to KEV - July 22. CVE-2026-50522 gets a July 25 federal remediation deadline.
Atoms Raises $1.7B - July 22. Travis Kalanick’s stealth industrial-AI robotics company, led by a16z, with Uber investing.
Monday.com Layoffs - July 22. About 630 jobs cut, 20 percent, in a restructuring around its AI Work Platform.
Uber Customer Service Cuts - July 22. 10 percent of the division let go, the first Uber layoff explicitly tied to AI.
OpenAI Voice for Codex - July 23. Full-duplex GPT-Live lands in the desktop app to drive Codex and ChatGPT Work hands-free.
Etched Raises $300M - July 23. Inference-chip startup hits a $10.3 billion valuation, the priciest Sequoia-led Series C ever.
Deno 2.9.4 - July 23. Point release upgrades V8 to 150.2.0 and improves Node.js compatibility.
Patreon Layoffs - July 23. 20 percent of staff, 93 people, cut in the platform’s largest reduction since 2022.
Claude Opus 5 - July 24. Anthropic’s new Opus model, near Fable 5 quality at half the price and unchanged $5/$25 pricing.
mrmustard PyPI Attack - July 24. Hijacked account ships an import-time credential stealer in the 0.7.4 release.

The week put the 2026 thesis in plain sight: intelligence is getting cheaper faster than anyone planned for, and the interesting product work has moved to routing, agents, and voice rather than raw model size. Anthropic priced frontier-adjacent quality at Opus rates, Google fragmented its lineup to compete on cost, and Cursor stopped making you choose a model at all. On the other side of the ledger, a SharePoint exploit and a poisoned PyPI package were fresh reminders that patched does not mean safe and that trusted release pipelines keep turning into malware delivery. Next week, watch for DeepSeek V4’s stable release, Moonshot’s Kimi K3 open weights on July 27, Node.js security releases the same day, and whether Google’s Gemini 3.5 Pro finally ships or slips again. See you then.