Top 10 AI News: July 10 2026 Daily Roundup

AI News Today July 10 2026: Top 10 Stories

Yesterday was the most consequential single day in AI model history. OpenAI launched GPT-5.6 Sol, Terra, and Luna publicly across ChatGPT, the API, and Codex on July 9, ending the 13-day government-coordinated preview. SpaceXAI launched Grok 4.5 the same morning, trained jointly with Cursor, priced at $2 per million input tokens and $6 output, and ranked fourth on Artificial Analysis's intelligence index. For the first time since the Fable 5 ban began on June 12, every major frontier AI lab has a publicly available model.

Today is Friday, July 10, 2026. SK Hynix begins trading on Nasdaq as SKHY, the largest ADR listing in history. Gemini 3.5 Pro reportedly has a July 17 date. The US Department of Health and Human Services just launched a ChatGPT audit program across all 50 states. And Google replaced its search results with an AI model. Here are the 10 stories every AI learner needs to know.

1. GPT-5.6 Sol, Terra, and Luna Go Public: What You Can Do Starting Today

GPT-5.6 Sol, Terra, and Luna became publicly available on July 9, 2026, across ChatGPT, the OpenAI API, and Codex. The 13-day government-coordinated preview that began June 26 with approximately 20 vetted partner organizations ended with OpenAI's announcement on X at 4:46 PM Pacific time on July 8: "GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday. We are expanding preview access globally now."

The rollout is staged rather than instantaneous. ChatGPT paid subscribers (Plus, Pro, Team, Enterprise) see GPT-5.6 tiers appearing in the model selector dropdown progressively throughout today and tomorrow. The full rollout is expected within 24 to 48 hours of launch. Free ChatGPT users retain GPT-5.5 as their default. The API model strings are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna, all accessible with valid API keys starting today. Codex receives all three tiers for the coding agent context.

Which Tier for Which Use Case

Terra ($2.50 input, $15 output per million tokens) is positioned as the default for most paid ChatGPT subscribers and the primary enterprise API tier. It delivers near-GPT-5.5 performance at half the cost and scores 84.3% on Terminal-Bench 2.1, tying Claude Fable 5. For the vast majority of business tasks, document processing, code review, customer support, and internal tooling, Terra is the right default.

Sol ($5 input, $30 output) is for the hardest agentic tasks where Terminal-Bench 2.1's 91.9% ultra mode score matters: long-horizon software engineering, complex multi-step reasoning, and security research workflows. Sol Ultra mode, which fans tasks to parallel sub-agents, carries an additional compute charge on top of Sol pricing. Luna ($1 input, $6 output) is for high-volume, latency-sensitive applications where cost dominates.

One important API migration note: the gpt-5.5-latest endpoint does not auto-migrate to GPT-5.6. You must explicitly update your model string to gpt-5.6-terra or gpt-5.6-sol to use the new models. Any production pipeline that does not update will continue running GPT-5.5, which is maintained as a supported API model.

My take: Terra is the model that changes enterprise AI economics this week. Near-GPT-5.5 performance at half the cost, from a model that also has the governance assurance that GPT-5.6 went through government pre-release review, is a compelling package for any enterprise that was routing to Chinese open-weight models for cost reasons during the past 13 days. The window to recapture those routing decisions is the next two weeks, before routing habits solidify into infrastructure.

2. Grok 4.5: SpaceXAI's First Post-Cursor Model at $2/$6 Per Million Tokens

SpaceXAI launched Grok 4.5 on July 8, 2026 for developers via Grok Build, Cursor, and the SpaceXAI console, with public rollout on grok.com and the X app on July 9. It is the first model released by SpaceXAI, the merged entity formed when SpaceX acquired Elon Musk's xAI in February 2026. It is also the first model jointly trained with Cursor, the AI coding editor SpaceX agreed to acquire for $60 billion in June.

The pricing is the most significant commercial element of the launch: $2.00 per million input tokens and $6.00 per million output tokens, with cached input at $0.50 per million. Output at $6 per million is roughly 4 times cheaper than Claude Opus 4.8 ($25 per million output) and identical to OpenAI's Luna tier. The context window is 500,000 tokens, smaller than Grok 4.3's 1 million token window and significantly smaller than Fable 5 or Opus 4.8's 1 million. SpaceXAI warns that requests above 200,000 tokens are billed at higher-context rates.

What the Cursor Training Actually Means

The joint training with Cursor is the technically interesting part of the Grok 4.5 story. Most AI coding models train on static code repositories, GitHub issues, and curated benchmarks. Grok 4.5 was trained on real developer session data from Cursor: debugging traces, multi-file diffs, user corrections, and the feedback signals that arise when a developer edits or rejects a suggestion. That is a qualitatively different training signal from static corpora.

Cursor disclosed one important caveat: an earlier Cursor codebase snapshot was accidentally included in training, which may give Grok 4.5 an advantage on CursorBench specifically. The benchmark is getting a larger update. Independent results should be compared on benchmarks other than CursorBench until that update is published. On benchmarks Cursor was not included in: Grok 4.5 scores 83.3% on Terminal-Bench 2.1 in standard mode and approximately 86% in agentic mode, below GPT-5.6 Sol's 91.9% but above Opus 4.8's 78.9%.

Grok 4.5 is not available in the EU at launch. SpaceXAI says EU availability is expected in mid-July, consistent with the company completing the regulatory notifications required under the EU AI Act for a new high-risk AI system.

My take: Grok 4.5 is a serious contender for high-volume agentic coding workloads where token efficiency matters more than absolute capability. The $6 output price is the story. At that rate, an agentic session that generates 100,000 output tokens costs $0.60. The same session on Fable 5 costs $5. On Opus 4.8 it costs $2.50. For teams running thousands of such sessions daily, the cost differential is structurally significant. The question is whether Grok 4.5's reliability holds up in production repositories as well as its benchmark performance suggests.

3. The Grok 4.5 vs Sol vs Claude Benchmark Reality Check

The simultaneous launches of GPT-5.6 Sol and Grok 4.5 on the same day have produced a wave of benchmark comparisons. Here is the most accurate picture based on published numbers from all sources, not just vendor marketing.

On Terminal-Bench 2.1: Sol Ultra scores 91.9% (the highest ever recorded), Sol standard scores 88.8%, Grok 4.5 agentic mode scores approximately 86%, Grok 4.5 standard scores 83.3%, Claude Mythos 5 scores 88.0%, Claude Fable 5 scores 84.3%, Claude Opus 4.8 scores 78.9%. Sol leads by a meaningful margin in standard mode. Grok 4.5 is between Fable 5 and Opus 4.8.

Where Grok 4.5 Actually Beats Claude

On DeepSWE 1.0, SpaceXAI's own published chart shows Grok 4.5 beating Claude Opus 4.8. On Terminal-Bench 2.1, it also beats Opus 4.8. On DeepSWE 1.1 and SWE-Bench Pro, xAI's own chart shows Opus 4.8 winning by 4 to 6 points respectively. The 'Opus-class' framing is accurate as a capability tier description. It is not the same as claiming to beat Opus across the board, which the published numbers do not support.

The most useful comparison is cost-adjusted performance. Grok 4.5 at $6 output versus Opus 4.8 at $25 output means that even if you are willing to accept 10 to 15 percent lower benchmark performance, you pay 76 percent less per output token. For tasks that require near-Opus performance but not full-Opus performance, Grok 4.5 is a legitimate routing target. Artificial Analysis ranked it fourth on its Intelligence Index, scoring 54, behind Fable 5, GPT-5.6 Sol, and Opus 4.8 but above GPT-5.5 and Sonnet 5.

On token efficiency, SpaceXAI reports that Grok 4.5 resolves SWE-Bench Pro tasks using an average of 15,954 output tokens versus 67,020 for Opus 4.8 (max). A 4.2x efficiency advantage is a genuine architectural difference, not a marketing claim. If accurate at production scale, it means Grok 4.5's effective cost per completed coding task is significantly below even its already-low nominal token price.

My take: Ignore the 'beats Opus' framing and look at the economics. Grok 4.5 offers roughly 75 to 80 percent of Opus 4.8's benchmark performance at roughly 24 percent of its output cost. For a significant class of production agentic coding tasks, that trade-off is the right one. Sol is still the best model for the hardest tasks. Grok 4.5 is now the best model for cost-sensitive high-volume agentic work. Sonnet 5 at $10 introductory output pricing sits between them economically, with its own agentic capability advantages.

4. SK Hynix Begins Trading Today on Nasdaq as SKHY

SK Hynix begins trading on the Nasdaq today, July 10, 2026, under the ticker SKHY. The $28 to $29 billion ADR offering at approximately $149 to $166 per ADS is the largest ADR listing in recorded market history, surpassing Alibaba's $21.8 billion New York debut in 2014.

The offering consists of 177.9 million American Depositary Shares, each representing one-tenth of an ordinary SK Hynix KOSPI share. Bank of America, Citigroup, Goldman Sachs, and JP Morgan are leading the offering. Cornerstone investors Baillie Gifford, Coatue Management, and Situational Awareness Partners have committed to purchasing up to $7 billion of the ADS.

SK Hynix holds approximately 60 percent of the global HBM market, delivering revenue of 52.6 trillion Korean won ($35.55 billion) in Q1 2026 alone, a 198 percent year-over-year increase. Operating margin reached 72 percent. The company's Korea-listed shares have surged more than 280 percent in 2026. HSBC forecast a 20 percent premium on the ADR listing, upgrading its Korea share price target 38 percent on the Nasdaq announcement.

NVIDIA and SK Hynix separately announced a multiyear technology partnership to co-develop next-generation memory for Vera Rubin AI supercomputers, Vera CPUs, RTX Spark PCs, and Jetson Thor robotics platforms. SK Hynix will also use NVIDIA's tools to accelerate semiconductor simulation and build digital twins for autonomous fab operations. The partnership formalizes what has been the de facto relationship between the two companies: SK Hynix supplies the HBM without which Nvidia's AI GPUs cannot function.

My take: Today is the day the AI memory trade goes retail in the US market. SKHY gives American investors frictionless exposure to the company that supplies 60 percent of the memory that runs every AI model they use. The risk is the boom-bust memory cycle. The bull case is that AI demand is structurally different from prior DRAM cycles because HBM complexity creates longer procurement windows and higher switching costs. Today's first-day trading will tell us how much of the 280 percent Korea rally investors are willing to pay for in the US market.

5. Gemini 3.5 Pro Leaked for July 17: Google Rebuilt the Pretraining from Scratch

Leaked details confirmed by multiple AI community sources place Gemini 3.5 Pro's general availability launch on July 17, 2026, eight days from today. The delay from the original June I/O commitment is now confirmed to have a specific cause: Google DeepMind abandoned the original 2.5 Pro base model and opted for completely new pretraining for Gemini 3.5 Pro, effectively rebuilding the model from scratch rather than fine-tuning or adapting the existing architecture.

New pretraining from scratch is not a minor correction. It means Google identified fundamental limitations in the 2.5 Pro foundation that could not be addressed through fine-tuning, RLHF, or post-training techniques. The decision to restart adds months to the delivery timeline but, if the new pretraining delivers the capability improvements Google is aiming for, it produces a qualitatively better model rather than an incrementally improved one.

The confirmed specifications remain: a 2-million-token context window (still the largest of any production frontier model by a factor of 2), Deep Think reasoning mode gated to the $250-per-month Ultra subscription tier, and pricing around $1.25 input and $10 output per million tokens for the standard tier. A new foundation model under Gemini 3.5 Pro, rather than an adaptation of Gemini 2.5 Pro, suggests the performance improvements over the 2.5 generation may be more significant than the prior roadmap implied.

My take: July 17 is a specific date. After three consecutive missed delivery commitments (May, June, and early July), a specific leaked date gives Google something to be accountable to publicly. If Gemini 3.5 Pro ships with new pretraining rather than a 2.5 Pro fine-tune, the resulting model may actually justify the wait. The 2-million-token context window combined with new pretraining optimized for long-context coherence could produce something that neither Sol nor Grok 4.5 can match at any price for large-codebase and large-document workloads.

6. HHS Deploys ChatGPT to Audit All 50 States for Fraud and Waste

The US Department of Health and Human Services announced it will use ChatGPT and other AI tools to analyze annual audit reports from all 50 states on an ongoing basis, targeting fraud, waste, and abuse in federal health spending. The program, led by Assistant Secretary Gustav Chiarello, has already alerted governors and treasurers in every state. HHS said the move addresses a longstanding gap where audit reports arrived but received little follow-up action.

Federal health programs, including Medicare and Medicaid, represent approximately $2.1 trillion in annual spending. Annual state audit reports covering those programs average several hundred pages each. Human analysts could not review all 50 reports comprehensively, let alone systematically flag patterns across states. ChatGPT enables HHS to ingest, analyze, and cross-reference all 50 reports simultaneously, looking for anomalies, inconsistencies, and known fraud patterns at a scale that was not operationally feasible before.

The announcement noted the program may result in federal funding being withheld from states that fail to correct identified deficiencies. That enforcement language elevates this from a research pilot to an operational compliance tool with real consequences. State governments that previously depended on delayed or insufficient HHS follow-up on audit findings are now dealing with an AI system that responds to every report consistently and comprehensively.

My take: The HHS ChatGPT audit program is the largest federal AI deployment for financial oversight ever announced. It is also the clearest example yet of the 'AI as force multiplier for small teams' use case. HHS does not need 50 teams of analysts to review 50 state audit reports. It needs one AI system that can do it consistently and flag what humans should review further. The enforcement consequence, potential funding withholding, makes this deployment consequential in ways most federal AI pilots are not.

7. Google Search Is Now Powered Entirely by Gemini 3.5 Flash

Google announced that its Search bar is now powered entirely by Gemini 3.5 Flash, generating custom AI-summarized pages in response to queries rather than traditional lists of links. Every search query on Google.com now returns an AI-generated summary page built by Gemini 3.5 Flash, with links to sources embedded within the AI summary rather than listed separately below it.

This is the most significant change to Google Search in its 27-year history. The traditional 10-blue-links format, which has been Google's core product since 1998, is being replaced by an AI-generated response that synthesizes information from multiple sources into a single document. For users who relied on skimming search result titles and URLs to find the right source to click, the new format requires reading the AI summary rather than the source list.

Google has been building toward this with AI Overviews, its AI-generated search summaries that have been expanding since 2024. The July 2026 announcement completes that transition: Gemini 3.5 Flash is no longer a supplementary feature added above search results but the primary interface for every query. Source links appear within the AI-generated page rather than as a separate list, which significantly changes how publishers receive traffic and how users discover content.

My take: Google replacing the search link list with a Gemini-generated page is the moment that changes the economics of web publishing permanently. Traffic from Google Search has been the primary distribution mechanism for most text content on the internet for 20 years. When Google generates its own AI page instead of linking to yours, your traffic disappears regardless of how good your content is. Cloudflare's decision to block AI training bots by default (announced last week) looks prescient in this context: publishers are simultaneously losing their Google distribution while AI companies harvest their content.

8. Sol on Cerebras: 750 Tokens Per Second Available Today

OpenAI's GPT-5.6 Sol is now available on Cerebras at up to 750 tokens per second, as announced alongside the general Sol launch. The Cerebras partnership, previewed in OpenAI's June 26 launch announcement, fulfills the promise of frontier-class AI at near-real-time speeds for developers who need interactive applications where response latency is the primary constraint.

To put 750 tokens per second in context: GPT-5.5 on standard API hardware delivers 30 to 80 tokens per second. A 1,000-token response that takes 12 to 25 seconds on standard infrastructure takes approximately 1.3 seconds at 750 tokens per second. The difference is not just speed perception. It changes the architecture of what AI applications are possible. Voice applications with no perceptible lag, code generation that completes before a developer loses their train of thought, and agent orchestration that runs multiple sequential steps within a single user interaction all become viable at 750 tokens per second in ways they are not at 50 tokens per second.

Cerebras' wafer-scale chip architecture achieves this by holding an entire large language model on a single die, eliminating the inter-chip communication latency that limits GPU cluster inference. The Cerebras deployment of Sol is available through the standard OpenAI API with a Cerebras routing option, at a speed premium above standard Sol pricing. Exact pricing for the Cerebras tier has not been separately disclosed.

My take: 750 tokens per second changes what AI can do, not just how fast it does what it already does. Latency at that level enables AI response times that feel instantaneous to humans, which opens voice, agentic, and interactive use cases that were architecturally constrained at typical GPU inference speeds. This is Cerebras's most significant commercial deployment to date and OpenAI's most important inference infrastructure announcement since the Jalapeño chip revealed with Broadcom last month.

9. SpaceXAI's Colossus Conflict: Training Competitors on Your Own Compute

Grok 4.5 was trained across tens of thousands of Nvidia GB300 GPUs, which SpaceXAI has at Colossus 1 and Colossus 2 in Memphis. Anthropic is paying approximately $1.25 billion per month for Colossus 1 access. Google is paying approximately $920 million per month for Colossus 2. Reflection AI is paying $150 million per month. Together, these three tenants are paying SpaceXAI roughly $2.3 billion per month to use the same compute infrastructure that SpaceXAI just used to train a model that directly competes with them.

This is not a legal problem. The compute lease agreements are straightforward commercial arrangements. SpaceXAI rents out capacity it has built. The tenants use that capacity for their own workloads. Nothing prevents SpaceXAI from using its own hardware for its own model training on the same infrastructure. But it creates an unusual dynamic: Anthropic is funding the infrastructure that trained its direct competitor's model. Every dollar Anthropic pays SpaceXAI in compute rent partially subsidizes the development of Grok 4.5.

Axios's Grok 4.5 launch article identified the structural tension directly: as SpaceXAI's own compute needs grow with each new model generation, it may have to choose between using capacity for its own models or leasing it to Anthropic, Google, and Reflection as a revenue stream. Grok 5, which is reportedly still training on Colossus 2 at approximately 1.5 gigawatts of power draw, will require even more compute. The revenue from Anthropic and Google is funding that training run. Whether SpaceXAI can continue to serve both roles indefinitely is the long-term question the Colossus tenant structure creates.

My take: The Colossus conflict of interest is the most interesting governance story in AI infrastructure that almost nobody is writing about directly. Anthropic and Google are contractually committed to paying SpaceXAI billions per month through 2029. SpaceXAI is using that contract revenue to train models that compete with them. Both parties understand this. Both parties have signed the contracts anyway. The rational explanation is that compute access at Colossus scale is so valuable that neither Anthropic nor Google has a better alternative, even knowing the conflict.

10. Illinois AI Safety Law: The First Frontier Model Regulation Signed in the US

Illinois Governor JB Pritzker signed SB 315 on July 6, 2026, making Illinois the first US state to sign into law frontier-model-specific AI safety requirements. Both OpenAI and Anthropic publicly supported the bill, which covers large AI developers meeting specific compute thresholds and creates four requirements: transparency obligations, catastrophic-risk assessment and mitigation, whistleblower protections for AI safety concerns, and independent oversight mechanisms.

The Illinois law's coverage threshold is based on compute: AI systems trained using more than 10 to the 26 floating point operations, a level reached by GPT-5.5-class and above models but not by smaller open-weight models. This compute-based threshold is a deliberate policy choice that focuses the law on the specific systems where the risk of catastrophic harm is most plausible, consistent with the UN Scientific Panel's assessment that no technical guarantee of AI safety currently exists for frontier systems.

The whistleblower protection provision is the most novel element. It gives AI company employees legal protection for reporting safety concerns to regulators without fear of retaliation, mirroring whistleblower protections in financial services and nuclear industries. The independent oversight mechanism requires covered AI systems to be subject to third-party safety audits, distinct from but complementary to the government-gated preview process under the federal June 2 Executive Order.

My take: Illinois passing the first frontier model safety law in the US is a significant step even though it covers only Illinois-based operations. The compute threshold approach is more legally defensible than capability-based definitions because compute is measurable and verifiable. The whistleblower protection is the provision most likely to have real effects: AI company employees who see internal safety corners being cut now have a legal structure to report concerns without risking their careers. That accountability mechanism has been missing from AI development entirely.

Frequently Asked Questions

Q: What happened in AI on July 9 and July 10, 2026?

July 9, 2026, is now documented as the most consequential single day in AI model history. OpenAI launched GPT-5.6 Sol, Terra, and Luna publicly across ChatGPT, the OpenAI API, and Codex, ending the 13-day government-coordinated preview. SpaceXAI launched Grok 4.5 the same morning at $2 per million input and $6 per million output tokens, trained jointly with Cursor. On July 10, SK Hynix begins trading on Nasdaq as SKHY, the largest ADR listing in history. Gemini 3.5 Pro has leaked a July 17 GA date.

Q: How do I access GPT-5.6 Sol, Terra, and Luna?

GPT-5.6 is available starting July 9, 2026. For ChatGPT: log into ChatGPT with a paid subscription (Plus, Pro, Team, or Enterprise) and check the model selector dropdown. Terra is expected as the default for standard paid subscribers. Sol is expected for Pro subscribers. Luna is available for high-volume or budget-oriented use cases. The rollout is staged and may take 24 to 48 hours to reach all accounts. For API access: use model strings gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna. The gpt-5.5-latest endpoint does not auto-migrate.

Q: What is Grok 4.5 and how does it compare to Claude Opus?

Grok 4.5 is SpaceXAI's new flagship model, launched July 8-9, 2026, trained on V9 foundation architecture and jointly trained with Cursor on real developer session data. It is priced at $2 input and $6 output per million tokens. It scores 83.3% on Terminal-Bench 2.1 in standard mode and approximately 86% in agentic mode. SpaceXAI's own benchmarks show Grok 4.5 beating Claude Opus 4.8 on DeepSWE 1.0 and Terminal-Bench 2.1, and losing on DeepSWE 1.1 and SWE-Bench Pro. Artificial Analysis ranks it fourth overall on its intelligence index, below Fable 5, Sol, and Opus 4.8. At $6 output vs Opus 4.8's $25 output, the cost advantage is the primary use case argument.

Q: How much does Grok 4.5 cost per million tokens?

Grok 4.5 is priced at $2.00 per million input tokens, $0.50 per million cached input tokens, and $6.00 per million output tokens. Requests above 200,000 tokens in the 500,000-token context window may trigger higher-context pricing. For comparison: Claude Opus 4.8 is $5 input and $25 output. Claude Sonnet 5 introductory is $2 input and $10 output. GPT-5.6 Luna is $1 input and $6 output. GPT-5.6 Sol is $5 input and $30 output. Grok 4.5 is approximately 4 times cheaper than Opus 4.8 on output and identical to GPT-5.6 Luna on output cost.

Q: When did SK Hynix start trading on Nasdaq?

SK Hynix begins trading on Nasdaq under ticker SKHY on July 10, 2026. The $28 to $29 billion ADR offering is the largest ADR listing in history, surpassing Alibaba's 2014 debut. Each ADS represents one-tenth of an ordinary SK Hynix KOSPI share, priced at approximately $149 to $166. SK Hynix holds 60 percent of the global HBM chip market and posted a 72 percent operating margin in Q1 2026 on $35.55 billion in revenue. NVIDIA and SK Hynix announced a multiyear partnership for next-generation AI memory on the same day.

Q: What is the Gemini 3.5 Pro July 17 release date?

Leaked details from multiple AI community sources place Gemini 3.5 Pro's general availability launch on July 17, 2026. The delay from Google's original June I/O commitment was caused by a fundamental decision to abandon the Gemini 2.5 Pro base model and rebuild Gemini 3.5 Pro from new pretraining. The model's confirmed specifications include a 2-million-token context window (the largest of any production frontier model), Deep Think reasoning gated to the $250-per-month Ultra tier, and pricing around $1.25 input and $10 output per million tokens.

Q: Is GPT-5.6 Terra now the ChatGPT default model?

GPT-5.6 Terra is expected to become the default for standard paid ChatGPT subscribers (Plus, Team, and Enterprise plans). The rollout is staged and may take 24 to 48 hours to reach all accounts from the July 9 launch. Free tier ChatGPT users retain GPT-5.5 as their default model. ChatGPT Pro subscribers have access to Sol. GPT-5.6 Luna is available as a selectable tier for high-volume or budget-constrained use cases.

Q: Did the US government use ChatGPT to audit all 50 states?

The US Department of Health and Human Services announced a program to use ChatGPT and other AI tools to analyze annual state audit reports for all 50 states on an ongoing basis, targeting fraud, waste, and abuse in federal health spending programs including Medicare and Medicaid. Led by Assistant Secretary Gustav Chiarello, the program has already alerted governors and treasurers in every state. HHS stated the program may result in federal funding being withheld from states that fail to correct identified deficiencies.