First Move  ·  AI Software  · 
Inside the labs — Tuesday morning, 23 June

Google DeepMind's Gemini 2.5 Pro Sets New Benchmarks; OpenAI Upgrades ChatGPT Amid Fable 5 Billing Shift

Google DeepMind just launched its most capable model, Gemini 2.5 Pro, while OpenAI rolled out a significant ChatGPT upgrade as Anthropic's Fable 5 free access period concluded.

ShareWhatsAppXLinkedIn

The story

The AI software landscape saw notable shifts this week as Google DeepMind unveiled its most capable model to date, Gemini 2.5 Pro with a new 'Deep Think' reasoning mode. This new model has reportedly set new highs across several benchmarks, including MMLU-Pro, GPQA Diamond, HumanEval+, and MATH-500, signaling a significant advance in model capabilities.

Meanwhile, OpenAI quietly shipped a major upgrade to ChatGPT over the weekend, with early reports suggesting it could be GPT-5.6, offering noticeable improvements in speed and capability. This comes as Anthropic's Fable 5 model transitioned to a metered credit system following the expiration of its free subscription window on June 22. Separately, the enforcement deadline for the EU AI Act is now just 41 days away, pushing major AI labs to accelerate compliance efforts, particularly around training data transparency, with potential fines reaching up to 7% of global annual revenue.

Who moved

Google DeepMind

What Changed: The lab launched Gemini 2.5 Pro, featuring a 2 million token context window and a 'Deep Think' reasoning mode.

Consequence: This new model achieved record-setting performance on key benchmarks, significantly advancing Google's position in the frontier model race.

OpenAI

What Changed: OpenAI shipped a major upgrade to ChatGPT over the weekend and launched its first formal global Partner Network with a $150 million investment.

Consequence: The ChatGPT upgrade hints at a more capable model, potentially GPT-5.6, while the Partner Network aims to certify 300,000 consultants to expand its ecosystem and enterprise deployment.

Anthropic

What Changed: The free subscription window for its Fable 5 model expired on June 22, with all subsequent usage now requiring metered credits.

Consequence: This change marks a significant shift in access and pricing for its advanced model, potentially increasing costs for power users.

DeepSeek

What Changed: DeepSeek unveiled its 1.6 trillion parameter V4 model, which was trained on Huawei chips.

Consequence: This move highlights the continued development of powerful models from Chinese labs and strategic partnerships in hardware.

Microsoft

What Changed: Microsoft announced the expansion of its Azure AI platform with five new industry-specific AI models.

Consequence: These models are designed to enhance productivity and compliance, accelerating enterprise AI adoption in sectors like healthcare, finance, and retail.

New models

Gemini 2.5 Pro with Deep Think

Lab: Google DeepMind

What: This model features a 2 million token context window and an extended inference process for complex problems.

Use: It is available immediately on the Gemini API, Google AI Studio, and Vertex AI, aiming to set new benchmarks for capability.

ChatGPT upgrade (likely GPT-5.6)

Lab: OpenAI

What: A significant, quiet upgrade to ChatGPT has been observed, with users reporting faster and more capable responses.

Use: This upgrade is expected to reclaim benchmark leadership for OpenAI, particularly in agentic coding tasks.

DeepSeek V4

Lab: DeepSeek

What: A 1.6 trillion parameter model trained on Huawei chips.

Use: This model's unveiling signals continued advancement in the capabilities of Chinese AI labs.

Muse Spark

Lab: Meta

What: Meta introduced a new AI model.

Use: Details are emerging, but it represents Meta's ongoing contributions to the AI model landscape.

Five new industry-specific AI models

Lab: Microsoft Azure AI

What: Models designed for specific sectors like healthcare, finance, and retail, incorporating domain-specific data and regulatory constraints.

Use: These models enable enterprises to deploy AI solutions faster and with greater accuracy within their industries.

Market signals

Venture capital investments in the first half of 2026 broke records, with OpenAI, Anthropic, and xAI collectively raising $237 billion.

Implication: This indicates a significant concentration of capital in a few frontier model companies, showcasing a 'power law' effect in AI investment.

The EU AI Act enforcement deadline is August 2, 2026, with non-compliance penalties potentially reaching up to 7% of global annual revenue.

Implication: Major AI labs are scrambling to meet transparency and disclosure requirements, which poses a substantial compliance challenge.

Anthropic's Fable 5 free subscription window expired on June 22, shifting all usage to metered credits at estimated rates of $10/$50 per million tokens.

Implication: This change represents a substantial increase in cost for power users, effectively doubling or tripling pricing for heavy coding sessions.

OpenAI launched a global Partner Network backed by a $150 million investment, aiming to certify 300,000 consultants by year-end.

Implication: This move signals OpenAI's strategic push to build out its implementation ecosystem and prepare for its anticipated public listing.

Builders saying

Demis Hassabis, CEO of Google DeepMind. He reaffirmed his prediction that artificial general intelligence (AGI) is only a few years away, potentially by 2030, calling it a profoundly transformative technology.

Yann LeCun, Chief AI Scientist at Meta. He reiterated his skepticism about AGI being close, stating that the concept of general intelligence is 'complete BS' and current transformer-based models are unlikely to achieve human-level intelligence.

What we'll be watching

Reporting + analyst voices: grounded via Google Search at publish time.