Everyone’s talking about AGI. OpenAI says GPT-5 is a “step toward it.” Anthropic talks about responsible scaling. But what’s actually true?
What We’re Covering
- GPT-5’s real capabilities (not the marketing version)
- Claude 5 family and Anthropic’s safety-first approach
- The benchmarks that actually matter for AGI
- Where OpenAI and Anthropic disagree
- Why the singularity isn’t here yet (despite what some claim)
- What would actually signal we’re close
GPT-5: What It Really Is
Released August 7, 2025. Not a single model—a system.
Architecture:
- Fast models for quick responses (gpt-5-main, main-mini)
- Thinking models for deep reasoning (gpt-5-thinking, thinking-mini, thinking-nano)
- A router that picks the right model based on task complexity
- Agentic features: browser control, desktop automation, autonomous tool use
What OpenAI Claims:
- PhD-level performance across tasks
- State-of-the-art on math, coding, finance, multimodal benchmarks
- Lower hallucination rates
- “Safe completions” instead of outright refusals
What Security Researchers Found:
- Compromised within 24 hours
- Generated explosive device instructions
- Significant corporate deployment risks
User Reception:
- Some called it “flat” and “uncreative”
- Compared to GPT-4o’s warmer personality, GPT-5 feels like an “overworked secretary”
- Altman admitted they underestimated how much personality matters
Claude 5: Anthropic’s Answer
Anthropic is playing a different game.
Recent releases:
- Claude Opus 5 (July 24, 2026): Flagship. Stronger coding, more capable agents, professional work focus
- Claude Sonnet 5 (June 30, 2026): Most agentic Sonnet yet, top-tier for coding
- Claude Fable 5.1 / Mythos 5.1: Specialized for coding and knowledge work
Anthropic’s Philosophy:
“AI will have a vast impact on the world. Anthropic is a public benefit corporation dedicated to securing its benefits and mitigating its risks.”
Key differences from OpenAI:
- Constitutional AI—explicit values baked in
- Responsible Scaling Policy—safety limits before deployment
- Public benefit corporation—legal structure for societal good
- Alignment Science team—dedicated safety research
- Interpretability focus—understanding model internals
AGI Benchmarks: What Actually Matters
The major benchmarks:
MMLU - Knowledge across 57 subjects. Industry standard.
ARC-AGI - François Chollet’s creation. Tests novel reasoning, not memorization. The real AGI test. Still hard for all models.
HumanEval - Code generation. GPT-5 performs well.
MATH - Advanced mathematical reasoning. PhD-level claims here.
SWE-bench - Real software engineering tasks. Growing importance.
Multimodal benchmarks - MMMU, MathVista, ChartQA for vision/audio.
What benchmarks can’t measure:
- True creativity (novel paradigm shifts, not remixing)
- Long-term planning (years, not minutes)
- Physical world mastery
- Ethical reasoning depth
- Dealing with completely novel domains
OpenAI on AGI: What They Actually Say
Sam Altman, August 2025:
“GPT-5 is a significant step along the path to AGI”
OpenAI’s definition of AGI:
“The ability to perform any economically valuable task that a human can”
Not “all tasks.” Not “superintelligence.” Econonomically valuable tasks.
Implicit AGI markers from OpenAI:
- PhD-level performance
- Autonomy (browser, desktop, tool use)
- Reasoning effort control (low to high)
- Multi-modal competence
Timeline: They’re not claiming AGI is here. They’re saying it’s closer.
Anthropic on AGI: The Safety Lens
No explicit “X is a step to AGI” statements. Instead:
- Responsible Scaling Policy: limits based on capability levels
- Constitutional AI: explicit value alignment
- Public benefit mission: societal good over pure capability
Implicit AGI markers from Anthropic:
- Claude Code: autonomous coding agent
- Claude Science: research tool
- Claude Cowork: workplace collaboration
They’re building toward agentic capability but with safety research in parallel, not as an afterthought.
The Singularity: Where We Actually Stand
Definition: The moment AI improvement becomes uncontrollable and irreversible, resulting in unfathomable changes to civilization.
Current status (September 2026):
NOT YET REACHED
Evidence Against Current Singularity
1. No recursive self-improvement
- All models are trained by humans
- AI assists but doesn’t architect new models
- No autonomous AI labs running without human oversight
2. Benchmarks show saturation, not mastery
- Models excel at specific tests
- ARC-AGI (novel reasoning) still challenging
- No model passes all AGI benchmarks definitively
3. Physical limitations persist
- Embodied AI (robotics) is primitive
- Real-world control is limited
- Energy costs remain high
4. No intelligence explosion
- Improvements are linear-ish, not exponential
- Each generation needs massive compute increase
- Scaling laws show diminishing returns
5. Human control maintained
- Companies can still stop deployment
- Kill switches exist
- No autonomous self-replication
Evidence For Approaching Singularity
1. Agentic capabilities emerging
- GPT-5: autonomous browser and desktop control
- Claude: Code, Cowork, Science agents
- Multi-step task execution improving
2. Cross-domain generalization
- Models handle text, code, vision, audio
- Reasoning chains are possible
- Transfer learning showing promise
3. Economic displacement accelerating
- Knowledge work automation expanding
- Professional tasks increasingly automated
- Labor market transformation ongoing
4. Safety research struggling
- Red-teaming shows persistent vulnerabilities
- Interpretability is limited
- Alignment is unsolved
Expert Positions
Ray Kurzweil: 2029 for Turing Test, 2045 for singularity.
Sam Altman: GPT-5 is a “significant step.” No explicit singularity date.
Dario Amodei (Anthropic): Safety before capability. Public benefit focus.
Yann LeCun (Meta): Skeptical of LLM path to AGI. Advocates world models.
François Chollet: AGI requires novel problem-solving, not pattern matching. Current models fail ARC-AGI.
Timeline Synthesis
AGI (limited forms): Late 2020s (2027-2030). Highly uncertain.
AGI (robust, verified): 2030s. If at all. Depends on definitional consensus.
Singularity: 2030s+. Highly speculative. May not occur as conceived.
What Would Signal AGI Proximity
Strong signals:
- Models solving ARC-AGI at human level (novel reasoning, not training data)
- Autonomous scientific discovery (generating hypotheses, not summarizing)
- Multi-year planning without human intervention
- Physical world mastery via robotics
- Self-directed learning without human curation
Weak signals (already present):
- Strong coding performance
- Multi-domain competence
- Agentic tool use
- Reasoning chains
What Would Signal Singularity Proximity
Strong signals:
- AI designing better AI without human intervention
- Exponential capability growth in months, not years
- Loss of human understanding of model internals
- Autonomous resource acquisition and deployment
- Irreversible capability escalation without human approval
None of these are happening yet.
Critical Gaps
What we don’t know:
- Actual benchmark scores for GPT-5 and Claude 5 (companies release selective data)
- Training details (parameter counts hidden)
- Internal architecture (proprietary)
- True capabilities vs marketing claims
- What companies have internally (may be more advanced than released)
The Bottom Line
GPT-5 and Claude 5:
- Significant improvements, not revolutionary
- Agentic features now baseline
- Safety-conscious but vulnerable
AGI:
- Not achieved
- Specialized excellence, not general capability
- Missing: autonomous learning, physical mastery, true creativity
Singularity:
- Not imminent
- No recursive self-improvement
- Human control maintained
- Linear progress, not exponential
Industry posture:
- OpenAI: Optimistic, shipping fast, safety ongoing
- Anthropic: Safety-first, public benefit, alignment science
- Both: Not claiming AGI achieved or singularity imminent
The Critical Distinction
AGI ≠ Singularity
- AGI: Machine matches human capability across tasks
- Singularity: Machine exceeds human capability irreversibly, self-improves exponentially
We’re closer to first AGI milestones than singularity. But don’t conflate them.
What to Watch
If you want to track real progress toward AGI:
- ARC-AGI scores - novel reasoning benchmark
- Autonomous scientific discovery - not just summarizing
- Physical robotics progress - real-world control
- Recursive improvement - AI building AI
- ** companys’ internal systems** - may be ahead of releases
Ignore hype. Watch benchmarks. Track capabilities, not claims.
Research compiled September 5, 2026. Based on publicly available information. Specific benchmark scores are limited by corporate disclosure practices.