The AI Wire ●

Anthropic launched Claude 3.7 Sonnet, its first hybrid reasoning model, with a standard mode for quick answers and an extended thinking mode that shows its step-by-step reasoning on hard problems. Users could control how long the model thought.

The launch arrived alongside Claude Code, Anthropic's agentic command-line coding tool, signaling a shift toward models that do extended work rather than just answer questions. Extended thinking quickly became the template for reasoning modes across the industry.

Claude 3.7 Sonnet bridged the Claude 3 and Claude 4 generations, keeping Anthropic competitive in coding while the next flagship was still in training.

xAI launched Grok-3, trained on its Colossus supercomputer with roughly ten times the compute of Grok-2. The release added 'Think' reasoning mode and DeepSearch, an agentic research tool that scours the web and X.

Musk called Grok-3 the smartest AI on Earth at launch, and early benchmarks put it at or near the frontier on math, science, and coding. The release event was streamed live with xAI engineers demoing the model.

Grok-3's compute-heavy approach validated xAI's bet on scaling infrastructure fast, and it became the base for the Grok-4 generation later in 2025.

Google launched Gemini 2.0 Flash, positioning it for what the company called the 'agentic era': native tool use, multimodal input and output, and the speed to power real-time agents. It was the first 2.0 model to reach developers.

Flash became Google's efficiency tier — near-flagship quality at a fraction of the cost and latency — a formula Google repeated aggressively through 2025 and 2026. The 2.0 generation also brought early agent prototypes like Project Astra and Project Mariner.

Gemini 2.0 Flash set the template Google followed ever since: lead with a fast, cheap Flash model, then bring the heavier Pro reasoning model behind it.

OpenAI released o1-preview, the first model in a new series trained with reinforcement learning to 'think' through problems using an internal chain of thought before answering. It was slower than GPT-4o but dramatically better at math, science, and competitive programming.

The launch introduced the reasoning-model paradigm that would define the next two years of AI development: trading inference-time compute for capability. A smaller o1-mini arrived the same day as a cheaper, faster alternative.

The full o1 followed on December 5 alongside the new $200/month ChatGPT Pro tier, cementing reasoning as a premium capability rather than a research curiosity.

xAI released Grok-2 in August 2024, its first frontier-class model, beating Claude 3.5 Sonnet on some benchmarks. It added image generation powered by Black Forest Labs' FLUX model, making Grok one of the first major chatbots to generate images natively.

The image generation launch was characteristically Musk: fewer content restrictions than rivals, which drew both users and criticism. Grok-2's benchmark gains showed xAI's Colossus training cluster paying off.

Grok-2 marked xAI's arrival as a genuine frontier lab rather than a fast follower, setting up the compute-heavy Grok-3 push in 2025.

Anthropic released Claude 3.5 Sonnet, a mid-tier model that outperformed the flagship Claude 3 Opus on many benchmarks while running at a fraction of the cost and twice the speed. It became the default model for most developers almost overnight.

The release also introduced Artifacts, a dedicated panel beside the chat where Claude could render documents, code, and interactive content. Artifacts became one of Claude's signature features and a template rivals later copied.

Claude 3.5 Sonnet's long reign — it stayed the developer favorite well into 2025 — made it arguably the most influential single model release of the Claude 3 generation.

OpenAI introduced GPT-4o — 'o' for omni — a single model natively handling text, vision, and audio with response times approaching human conversation. It was faster and cheaper than GPT-4 Turbo, and unusually, it rolled out to free ChatGPT users too, with usage limits.

The headline feature was Advanced Voice Mode, letting users talk to ChatGPT nearly in real time with interruptions and emotional tone. It began rolling out to Plus users in late July and became one of the product's signature experiences.

GPT-4o remained the flagship for over a year and developed an unusually loyal following — its eventual removal from ChatGPT years later prompted genuine user backlash over its personality.

xAI released Grok-1.5 with a 128,000-token context window and substantially stronger reasoning, closing the gap with frontier rivals. A vision variant, Grok-1.5V, added image understanding on April 12.

The release showed xAI could iterate quickly: five months after the Grok debut, the model was competitive on standard benchmarks. Musk promised Grok-1.5 would reach X Premium+ subscribers within days.

Grok-1.5 set up the bigger Grok-2 push later in 2024, when xAI aimed to move from catching up to leading.

Anthropic unveiled the Claude 3 model family with three tiers for the first time: the lightweight Haiku, the balanced Sonnet, and the flagship Opus. Opus and Sonnet launched on March 4, with Haiku following on March 13. The family added native vision, handling images alongside text.

Anthropic said Claude 3 Opus beat GPT-4 and Gemini 1.0 Ultra on several common benchmarks, the first time a Claude model claimed the outright frontier lead. The three-tier structure let developers trade cost, speed, and intelligence per task.

Claude 3 cemented the family-and-tiers naming that Anthropic kept through every subsequent generation, and made multimodality standard across the lineup.

Google released Gemini 1.5 with a 1-million-token context window — an order of magnitude beyond rivals at the time — later extended to 2 million tokens. It could process hour-long videos, huge codebases, or entire books in a single prompt.

The release used a Mixture-of-Experts architecture for efficiency, delivering flagship-class quality at lower serving cost. Google pitched long context as the defining advantage of the Gemini line.

Gemini 1.5's context breakthrough forced the whole industry to compete on long context, a race that continued through the 2.x and 3.x generations.

Google unveiled Gemini 1.0 in three sizes — Ultra, Pro, and Nano — calling it the company's first natively multimodal model, trained on text, images, audio, and video from the start. Gemini Pro powered Bard immediately; Ultra rolled out later after further tuning.

Google claimed Gemini Ultra was the first model to outperform human experts on the MMLU benchmark suite, a milestone it used to declare the frontier lead. The launch was Google's full-scale entry into the flagship race after the Bard stopgap.

Gemini 1.0 set the naming and the multimodal-first positioning Google has kept through every generation since, including the Nano tier that later shipped on Pixel phones.

At its first developer conference, OpenAI announced GPT-4 Turbo: a faster, cheaper GPT-4 with a 128K-token context window — roughly 300 pages of text — and knowledge updated to April 2023. API pricing dropped significantly, a direct response to developer complaints about GPT-4's cost.

DevDay also introduced GPTs — custom chatbots anyone could build without code — and the GPT Store concept. The message was clear: OpenAI wanted an app ecosystem on top of its models, not just a chatbot.

The event was overshadowed days later by the board's brief firing of Sam Altman, but GPT-4 Turbo itself became the workhorse model of the GPT-4 era, powering ChatGPT Plus for over a year.

xAI debuted Grok in November 2023, a chatbot for X Premium+ subscribers pitched as witty and rebellious, with real-time access to posts on X. It was Elon Musk's answer to ChatGPT, built by his new AI company barely four months after its founding.

Grok's X integration gave it a genuine differentiator: answers informed by live social media. In March 2024 xAI open-sourced Grok-1's weights, a surprise move that won goodwill with the open-source community.

The launch established xAI's playbook — move fast, differentiate on X data and personality, and iterate the Grok line aggressively.

Anthropic launched Claude 2, its first model widely available to the public through a beta at claude.ai and the API. Its headline feature was a 100,000-token context window, which could swallow entire books or long documents in a single prompt.

Claude 2 was the first model widely seen as a genuine ChatGPT rival, with Anthropic claiming improvements in coding, math, and reasoning over its predecessor. The public beta expanded access beyond the UK and US rollout that followed.

The release set the pattern for Anthropic's strategy: longer context and stronger reasoning pitched as the responsible alternative to OpenAI's faster-moving lineup.

OpenAI launched GPT-4, its first flagship successor to the model behind ChatGPT. It was multimodal from the start — accepting image inputs alongside text — and dramatically stronger at reasoning, instruction-following, and exams, scoring in the top 10% on a simulated bar exam.

GPT-4 debuted for ChatGPT Plus subscribers and as a waitlisted API, with 8K and later 32K context variants. Microsoft confirmed its new Bing chat was already running on GPT-4, making it the quiet engine of the search industry's AI pivot.

The release also marked a shift in OpenAI's openness: the company published a system card and red-teaming results but disclosed almost nothing about architecture, training data, or model size — a stance that drew criticism from researchers.