Google drops three new Gemini Flash models but skips the Pro update everyone was waiting for. Bitcoin is stuck in a $2,500 range with a critical Sunday close ahead, and CME futures open interest just hit its lowest level since October 2023. Metaplanet's subsidiary raises nearly $60 million to keep stacking sats despite a brutal year for its stock. And in Washington, the CLARITY Act has four working days to survive. Let's get into it.
Google shipped three new Gemini models this week, and the interesting thing is what's not in the lineup. No Gemini 3.5 Pro. That's still in partner testing, and there's already a pre-training run going for Gemini 4. Instead we got 3.6 Flash, 3.5 Flash-Lite, and a niche one called 3.5 Flash Cyber.
Start with 3.6 Flash. This is the workhorse. Google claims it uses 17% fewer output tokens than 3.5 Flash on comparable tasks, and in some coding benchmarks up to 65% fewer. Output token pricing drops from $9 to $7.50 per million, input stays at $1.50. On DeepSWE it jumps from 37% to 49%. On MLE Bench, from about 50% to nearly 64%. Translation: fewer wasted edits, fewer reasoning loops, cheaper agent runs. That last part matters more than the benchmark numbers. If you're running agents in production, token efficiency is the whole game.
Then there's 3.5 Flash-Lite, which cranks out 350 output tokens per second at $0.30 input and $2.50 output per million tokens. It ships with a built-in computer-use tool. Google is very clearly aiming this at high-volume agentic search and document processing workloads.
And 3.5 Flash Cyber is the one worth watching. It's fine-tuned for finding and patching vulnerabilities, paired with an agent framework called CodeMender. In V8 engine testing it flagged 55 unique confirmed issues versus 47 for plain 3.5 Flash and 36 for Claude Opus 4.6. Google's Big Sleep team says it's already catching hard-to-find bugs in Chrome, Android, and internal Cloud codebases. One team reportedly found a remote code execution vulnerability in two hours with a 100% reliable exploit bypassing ASLR. Access is limited, gated to governments and trusted partners, for obvious reasons.
The strategic read: Google isn't trying to win the frontier benchmark race this month. They're carpet-bombing the middle of the market where agents actually run. Cheap, fast, specialized. Meanwhile OpenAI and Anthropic keep shipping. Everyone's iterating.
On the same theme, agentic coding is quietly crossing a line from assistant to autonomous worker. A few things landed this week that make that concrete.
GitHub Copilot's Cloud Agent is now generally available inside Linear. You assign a Linear ticket to Copilot, it spins up an isolated GitHub Actions environment, reads the codebase, drafts a solution, and opens a pull request. Progress updates stream back into the ticket. You can @ mention it with follow-up instructions. It's not writing production code unsupervised, but the loop from ticket to draft PR is now measured in minutes without a human touching the keyboard.
Claude Opus 5 also landed in Copilot this week for the Pro Plus, Max, Business, and Enterprise tiers. Anthropic added stronger safeguards around cyber content, which is worth noting given what Google is doing on the offensive side with Flash Cyber. Different philosophies, same underlying tension.
OpenAI went a different direction. They pushed GPT-Live's full-duplex voice into the ChatGPT desktop app on macOS and Windows, wired into Codex and ChatGPT Work. You can now talk to it while it runs multi-threaded coding tasks in the background, orchestrating debugging, PR reviews, and unit test generation across Slack, GitHub, and local repos from a single spoken prompt. The heavy reasoning gets offloaded to GPT-5.5 in the background while the voice model handles the conversation.
And JetBrains shipped Rider and ReSharper 2026.2, embedding AI agents with actual project context: coverage data, profiler output, the refactoring engine. ReSharper introduced Junie, an LLM-agnostic agent that can run terminal commands, manage git, and refactor autonomously inside Visual Studio.
The direction here is unmistakable. The IDE is no longer where you write code. It's where you supervise agents that write code. Whether that's good for junior developers learning the craft is a different conversation.
Metaplanet keeps writing the playbook for non-US Bitcoin treasury companies, and this week they added a new chapter. Their Tokyo-listed subsidiary, called Bitcoin Japan, closed a financing arrangement with EVO Fund worth up to ¥9.66 billion, roughly $59.5 million. It's a mix of zero-coupon convertible bonds and stock acquisition rights. The first ¥662 million, about $4 million, is earmarked for an immediate Bitcoin purchase. The rest funds broader expansion.
What's interesting is the structural move. Metaplanet's parent already holds around 43,000 BTC valued near $2.8 billion. But instead of just buying more on the parent balance sheet, they're building a subsidiary-level financing vehicle that can issue its own instruments and stack independently. That's a maturing treasury architecture, one that lets them scale without constantly diluting the parent.
Context matters here. Metaplanet stock is down more than 50% year-to-date. The parent recently revised its target to 210,000 BTC by end of 2027, which at current prices would require tens of billions in capital. So this subsidiary financing isn't a victory lap. It's them doubling down while their equity is in the tank.
Meanwhile in the US, Strive, formerly Semler-plus-Asset-Entities, added another 21 BTC at an average of about $63,221, bringing them to 19,921 BTC with $157 million in cash. Small buy, but they're funding it through perpetual preferred equity rather than debt. Strive posted a $393 million net loss over its first six months as a public company, mostly from bitcoin fair-value accounting and share issuance treatment. Management is openly saying they want a bigger war chest.
Zoom out. Galaxy notes that dormant Bitcoin movement just hit a four-year low. OGs have slowed distribution. If long-term holders are done selling and treasury companies are still buying with structured debt, the float keeps tightening. That's the setup, regardless of what happens on this week's chart.
Bitcoin closes the week near $64,500, wedged between $65,000 resistance overhead and $62,500 support below. It touched $66,990 on July 21, a one-month high, then slipped back and turned that breakout level into a ceiling. Sunday's close matters. A reclaim of $65k opens the door toward $68k. A rejection puts $62,500 in play, and below that, $60,000.
But the more interesting story is what's happening in the derivatives stack. CME Bitcoin futures open interest just fell to 90,030 BTC after the June expiry. That's the lowest reading since October 2023, down about 58% from the December 2024 peak. Front-month contango is essentially zero, annualized basis under 4%, dipping near 2% in illiquid hours. Institutions are deleveraging, hard.
At the same time, perpetual swap funding rose to about 6.25% annualized, and perp open interest climbed by 15,600 BTC while CME open interest dropped 8,000 BTC. So the leverage isn't gone. It migrated. From regulated US venues to offshore perpetuals. Retail and crypto-native funds are running the tape while institutions sit out.
Options tell a similar story. The biggest bullish position on Deribit shifted from the $80,000 call down to the $70,000 call, with $1.63 billion in open interest there. Max pain across Deribit, CME, Binance, and OKX has clustered around $60k to $65k, basically where spot sits. One notable trade over the July 17-19 weekend was a bull call spread, long the $70k call and short the $72k call, both July 31 expiry. Traders are positioning for a modest, targeted upside, not a moonshot.
One more angle. Miners get a roughly 16% difficulty drop around today, July 26. That's a real lifeline for the machines still online. But $19 billion in AI compute deals is pulling the largest miners toward HPC hosting instead. The difficulty adjustment helps whoever stays committed to hashing. Everyone else is being courted by hyperscalers with better contracts.
When institutions unwind CME leverage to a two-year low while dormant coins refuse to move and treasury companies keep buying with structured debt, that's not a bearish setup. That's a coiled one.