xAI Releases Grok 4.3 with Focus on Agentic Tasks
Elon Musk's new model targets practical efficiency and task execution, offering a cheaper alternative.
The news about Grok 4.3 is presented by the media as a technical announcement: the model has improved on agentic tasks and become significantly cheaper. But I see here a story not about artificial intelligence, but about Elon Musk's desperate attempt to rebuild the business model of his entire tech conglomerate amid the personnel and financial collapse of xAI.
The Core: What's Really Happening
The launch of Grok 4.3 occurred on April 30, 2026, and was accompanied by an uncharacteristic silence from Musk. No flashy presentations, pompous statements about "conquering the universe," or promises of imminent AGI. Instead, a dry update on the website with API documentation. This is a radical shift in strategy, driven by the desperate state of affairs inside the company.
By spring 2026, xAI had become a personnel disaster. All 11 co-founders had left the company amid an internal assessment of its technological lag as "humiliatingly low." The utilization rate of the Colossus supercomputer with 22,000 GPUs was only 11%, which, with monthly operating costs of around $10 billion, meant massive losses—compare this to the industry average utilization rate of 43-46%. In effect, Musk was left with a huge and expensive fleet of equipment, a team of newcomers, and a product that was losing to competitors on all key benchmarks.
It is in this context that Grok 4.3 emerged. This is not just a model update, but an urgent shift in direction from chasing leadership to capturing the price-performance niche. Key figures: input cost dropped to $1.25 per million tokens, and output to $2.50. That's a reduction of nearly 60% compared to its predecessor. For comparison, direct competitor Claude Opus 4.7 from Anthropic charges $5 for input and $25 for output. Grok 4.3 is 10 times cheaper.
Timeline and Context
The release of Grok 4.3 cannot be viewed in isolation from parallel events. On April 30, 2026, simultaneously with the model's release, Musk began forcibly phasing out the entire previous product line. Eight older models, including Grok 4-fast, Grok 3, and the specialized grok-code-fast-1, will be fully shut down on May 15, 2026. Developers have only 9 working days to migrate. This is an unprecedentedly short timeframe, indicating not a planned transition but an emergency one, driven by the need to urgently cut costs for supporting outdated infrastructure.
At the same time, on May 1, 2026, immediate availability of Grok 4.3 on the Oracle Cloud Infrastructure (OCI) platform was announced. This is a key point. Musk traditionally built his own data centers, but now he has entered a strategic partnership with Oracle, effectively admitting xAI's inability to handle scaling on its own. The calculation is simple: rapid access to Oracle's corporate clients with their multi-billion dollar AI budgets. At stake are not contracts with retail users at $30 per month, but enterprise deals where the annual check can reach $500,000 or more.
Who Wins and Who Loses
Winners:
- Oracle. Larry Ellison gained an exclusive partner with a model that can compete on price with Chinese solutions, and now sells it as part of the OCI Enterprise AI package. This gives Oracle an edge in the battle against AWS (Anthropic) and Microsoft (OpenAI).
- Budget-constrained enterprise clients. Companies already spending six-figure sums monthly on Anthropic or OpenAI APIs get direct savings. One test engineer reports that running 18 complex agentic tasks on Grok 4.3 cost him $7.84, while the same tasks on Opus 4.7 cost $71.50. That's nearly a 10x difference. For a startup making thousands of API calls per day, switching to Grok means saving tens of thousands of dollars monthly.
- Musk himself. More precisely, his project to take SpaceX public. Too few noticed, but Musk has already mandated that all banks and law firms involved in preparing the IPO purchase corporate subscriptions to Grok. This creates guaranteed initial demand and a revenue pipeline.
Losers:
- Developers whose products depend on older xAI models. Nine days to migrate is an unreasonably short timeframe that will break many services on May 15.
- Free-tier users. xAI is openly scaling back free access. The most powerful features of Grok 4.3, including custom voices and agentic tools, require a SuperGrok subscription at $30/month or X Premium+ at $40/month. This is the end of the "Grok for everyone" era; Musk is focusing on monetization.
- OpenAI. Although their models still lead in quality, aggressive price dumping by xAI will pressure their margins. GPT-5.5 costs clients $35 per million tokens (input+output) versus $3.75 for Grok 4.3. Retaining corporate clients with such a price gap is only possible through unquestioned leadership in complex reasoning.
What the Media Isn't Saying
Here's an insight I reconstruct from the mosaic of facts: Grok 4.3 is a model created not to conquer the market, but to make xAI an attractive asset for sale. Musk sees that the battle of foundation models is lost: GPT-5.5 and Claude 4.7 Opus are objectively stronger in complex cognitive tasks. But they are expensive. Musk has created a "budget killer"—a model good enough for 90% of business tasks (agentic chains, report generation, working with PDF and Excel) and anomalously cheap.
The second non-obvious point concerns the new monetization model for "thinking." Grok 4.3 has a built-in and non-disablable reasoning mode, Always-on reasoning. This means the model always "thinks" before answering, generating so-called reasoning tokens. And they are also billed to the user at the full rate. So the stated $2.50 per million output tokens is just the base rate. The actual cost of a complex query can be significantly higher due to hidden generation of internal reasoning chains. It's like paying a taxi driver not only for the miles traveled but also for the miles he drove while thinking about the route before you got in the car. Moreover, xAI has introduced a unique market fee of $0.05 per request blocked by the safety system before processing even begins. You pay even for being denied an answer.
Forecast: Next 30 Days and 90 Days
30 days (by June 10, 2026):
We will see an exodus of developers from older xAI models. The May 15 deadline will turn into chaos for those who haven't migrated. Then, by early June, a wave of startups lured by the low price will begin moving their agentic pipelines from Claude and GPT to Grok 4.3. This won't be a mass exodus, but rather a segmentation: routine, high-volume tasks will go to Grok, while complex expert tasks will remain with the leaders.
90 days (by August 2026):
Musk will begin negotiations to merge or sell xAI. The buyer could be Oracle, which has already deeply integrated the model into its ecosystem. The deal price, in my estimation, will range from $15 to $22 billion—roughly half of xAI's peak valuation. An alternative scenario is xAI being absorbed by X Corp, with a formal explanation about "synergy between social network and AI," but in reality to hide a financial hole within the larger structure before the IPO. In any case, Grok 4.3 will go down in history not as a technological masterpiece, but as the model that brought to the AI market something it hadn't seen before—a real price war.
— Editorial Team
No comments yet.