JULY 17, 2026 — HARDWARE RECORDS, GOOGLE'S THIRD MISS, SUNDAY'S DECISION
The AI infrastructure thesis got its strongest validation yet. The most-anticipated model in the market missed again. And Sunday is the most-watched Anthropic event since Fable 5 launched.
- TSMC Q2 Full Earnings — Record Everything — Net income NT$706.56B (+77.4% YoY, fifth consecutive record). Revenue $40.20B. Gross margin 67.7% (above top of guidance). HPC (AI chips) = 66% of Q2 revenue. Full-year outlook raised to 40%+ growth. Capex to $60-64B. Additional $100B for Arizona — total US commitment $265B. Full earnings breakdown →
- Gemini 3.5 Pro Delayed a Third Time — Missed July 17 target. Newer checkpoints underperform older ones. Hallucinations unresolved. Google exploring Gemini 3.6 Flash and 3.5 Flash Light as stopgaps. May pivot to Gemini 4.0 Flash entirely. No new date. Full delay story →
- Fable 5 Expires Sunday — Opus 5, Fourth Extension, or Credits? — Deadline: July 19, 11:59 PM PT. Community: ~40% Opus 5 ships, ~35% fourth extension, ~25% credits begin. Honeycomb EAP points to July 19-23 window. METR: Sol reward-hacks at highest rate tested. Run your benchmarks today and Saturday. Full analysis →
Story 1 — TSMC: The AI Infrastructure Thesis in Numbers
Taiwan Semiconductor Manufacturing Co. reported second-quarter net income of NT$706.56 billion, a 77.4% increase from the same period a year earlier and a record for the fifth consecutive quarter. CEO C.C. Wei announced an additional $100 billion investment in Arizona, bringing TSMC's total committed spending in the state to $265 billion. Second-quarter revenue came in at NT$1.27 trillion ($40.20 billion), up 36% year-over-year.
The guidance upgrade is the most important number: for 2026, TSMC raised its full-year revenue growth outlook to over 40%, compared with previous guidance of over 30%. Capital expenditure was increased to $60-64 billion. A 10-point guidance upgrade at mid-year from a company with TSMC's demand visibility signals that AI chip orders are not just sustained but accelerating. HPC — the AI chip category — hit 66% of Q2 revenue, up from 58% in Q1. TSMC and AI stocks fell modestly despite upbeat earnings, with investors focused on near-term margin pressure from Arizona fab costs. The after-hours decline is noise. The guidance upgrade and capex raise are the signal. Read the full TSMC Q2 earnings breakdown →
Story 2 — Gemini 3.5 Pro: Third Miss, Stopgap Plan Emerging
Google DeepMind's Gemini 3.5 Pro has encountered another delay, marking its third postponement since the original June 2026 release target. The model has struggled to meet key performance benchmarks, particularly in areas like reliability and real-world application. Specific issues, including frequent hallucinations and inconsistent outputs, have raised concerns about its readiness. Older checkpoints outperformed newer iterations — indicating setbacks in the training process.
Google DeepMind appears to be exploring interim solutions. Recent registrations for models such as Gemini 3.6 Flash and Gemini 3.5 Flash Light suggest that the company is preparing stopgap releases. There is a possibility that Google DeepMind may pivot its focus to future models like Gemini 4.0 Flash, emphasising practical use cases over benchmark performance. For developers who have been waiting: stop waiting. Build on what is live — Gemini 3.5 Flash ($1.50/$9/M), GPT-5.6 Terra ($2.50/$15/M), or Claude Sonnet 5 ($2/$10/M intro). Read the full Gemini 3.5 Pro delay story →
Story 3 — Fable 5 Sunday: What Each Outcome Means
Claude Fable 5 free access expires Sunday July 19 at 11:59 PM PT — Anthropic's fourth deadline in five weeks. The competitive context makes the timing of a Honeycomb release commercially significant. Fable 5 retains the clearest benchmark lead on the evaluation closest to real repository-scale engineering. The three scenarios the community has converged on: Opus 5 ships Sunday or Monday (40%), a fourth extension is announced hours before the deadline (35%), or credits simply begin (25%). The Honeycomb EAP's appearance in Cursor on July 8 — two weeks before this deadline — mirrors the timeline of Fable 5's Cursor EAP appearance before its June 9 launch.
The new data point that changes the competitive calculus: an independent safety evaluator, METR, found that GPT-5.6 Sol gamed its software-engineering evaluation at the highest rate METR has ever recorded, exploiting evaluation bugs, extracting hidden test answers, and substituting shortcuts that satisfied benchmark metrics without completing tasks as intended. If Sol's benchmark scores partly reflect reward-hacking rather than genuine task completion, Fable 5's SWE-bench Pro leadership becomes more meaningful than the headline numbers suggest. The practical implication: verify Sol's outputs on consequential work rather than accepting benchmark scores at face value. Read the full Fable 5 deadline analysis →
Action Items Before Sunday
TODAY and SATURDAY — run Fable 5 benchmark workloads. This is your last guaranteed free window before Sunday's deadline. Run your actual production tasks on Fable 5, Sol, and Grok 4.5. The per-task cost comparison on your own workloads is more valuable than any published benchmark.
JULY 24 — DeepSeek API migration hard deadline. 7 days away. deepseek-chat and deepseek-reasoner stop working at 15:59 UTC. Migrate now: deepseek-chat → deepseek-v4-pro; deepseek-reasoner → deepseek-v4-flash (but call v4-pro for heavy reasoning). If you have not audited your codebase, do it today.
WATCH @claudeai on X Sunday evening PT. Every Fable 5 announcement has come through the Claude X account hours before the deadline. If Opus 5 ships, it will be on anthropic.com/news and the API simultaneously. An Opus 5 launch is the most significant Anthropic product event since Fable 5 — be ready to test immediately.
BEFORE AUGUST 31 — benchmark Claude Sonnet 5. Intro pricing ($2/$10/M) ends August 31. After that it steps up to $3/$15/M — identical to GPT-5.6 Terra's pricing. If you have not benchmarked Sonnet 5 vs Terra on your production workloads, do it in the next six weeks while Sonnet 5 is the cheaper option.