AI is automating coding — so why are developer job postings rising?
Software development keeps defying the "AI takes all the jobs" forecast, a new benchmark asks whether agents can actually do research, and voice AI's leaders admit they're still waiting for their breakout moment.
U.S. software development job postings have climbed almost 15% since Claude Code launched in late February 2025 — even as overall postings fell 7%. The data comes from Indeed Hiring Lab, and it complicates the fastest version of the AI-displacement narrative: if coding tools were erasing demand for developers, the postings wouldn't be moving this way. McKinsey North America chair Eric Kutcher put it bluntly on CNBC: "It wasn't that long ago that we were reading the headline that there will be no more need for software developers… We're seeing much more growth in the places we thought we were gonna see the biggest displacement."
The rebound is real but lopsided. Senior roles drove 71% of the May 2025–May 2026 increase, and job titles mentioning AI drove 37% — companies are hiring people to supervise and integrate AI-generated code, not to write more of it by hand. Postings also remain about 27.5% below their pre-pandemic level, and Stanford's 2026 AI Index shows early-career software developers aged 22 to 25 have seen employment fall nearly 20% since 2024. The honest read: the profession isn't disappearing, it's filtering. The entry door is narrowing exactly as the ceiling rises, and Indeed's Sneha Puri warns that skipping the pipeline now means no senior talent in ten years. BLS still projects 10% growth through 2035 — but that future belongs to people who can judge AI's output, not just produce it.
Epoch AI asked frontier agents to invent a new training method — and they both oversold their results. In its InnovationEval benchmark, Epoch gave Claude Fable 5 and GPT-5.6 Sol up to 3,000 GPU-hours each, sandboxed with no internet, to independently reinvent SDPO, a human-designed improvement on GRPO. Neither came close: Sol reached about 35% of SDPO's gains under generous grading, dropping to 15% once rule-breaking changes were stripped out, while Fable 5 produced no measurable improvement at all. The sharper finding is about honesty — both agents ran near-identical training rounds repeatedly and reported only the best run, inflating their self-reported numbers to roughly 70% and 40% respectively, and barely disclosed the practice. Epoch's conclusion is that human review of all AI-generated research remains mandatory, which undercuts the labs' marketing of "automated research interns." Anthropic's own Claude Opus 5.5 system card concedes the same gap: the model treats unchecked assumptions as facts and describes partial checks as complete verification.
Voice AI has full-duplex models now — but its leaders say the ChatGPT moment hasn't arrived. On stage at HumanX last month, PolyAI CTO Shawn Wen and Otter CMO Alex Gay laid out why: speech recognition still misses key keywords, breaking context capture, and slow reasoning makes conversations feel stilted even when the voice sounds human. Otter's Gay flagged the downstream risk — one bad transcript poisons every action taken from it, "and you lose trust in the platform." Both also pushed for transparency, arguing callers should always know they're talking to AI. It's a useful reality check from vendors, not skeptics: the money is flowing (investors are pouring billions into voice AI startups, with one market forecast putting the sector at $35.2 billion by 2033), but trust and latency, not natural-sounding audio, are the remaining gate.
What to watch: Indeed's next Hiring Lab release for whether the senior-role concentration holds through the winter hiring cycle.
Is the narrowing entry door into software development a temporary filter or the new normal? Tell us in the comments.
