Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
OpenAI shipped GPT-5.6 to the general public on July 9, 2026. The release closed out a five-week detour that started with a presidential executive order and ended with the company’s most capable model sitting in front of a US government review board before anyone outside a short list of vetted firms could touch it. Now the gate is open. Sol, Terra, and Luna, the three tiers that make up the GPT-5.6 family, are live across ChatGPT, the API, and Codex for anyone with an account.
The launch matters beyond the spec sheet. It marks the first time a frontier AI lab has pushed a flagship model through a mandatory pre-release government check, and it lands at a moment when OpenAI’s hold on the AI chatbot market is loosening on several fronts at once, from Anthropic’s enterprise pull to Google’s default placement inside Search. This is a follow-up to Tech Insider’s earlier coverage of the gated rollout, and it fills in what happened after the waiting period OpenAI promised.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
On June 26, 2026, OpenAI made GPT-5.6 available to roughly 20 organizations, each one individually vetted by federal officials before access was granted. That arrangement traced back to an AI cybersecurity executive order signed in early June, which required labs to submit their most capable models for government review 30 days ahead of public release. Thirteen days after the gated preview began, the review cleared and OpenAI opened Sol, Terra, and Luna to every ChatGPT, Codex, and API user, according to Engadget’s reporting on the July 9 launch.
OpenAI pushed back on the process even as it complied. The company said it does not believe this kind of government access arrangement should become the long-term default, while acknowledging that cooperating with the review helped it reach the public faster than a drawn-out standoff would have. TechCrunch’s original report on the restriction and a parallel account from Forbes both describe the same tension: a lab racing to ship its best model while a government office decides who gets to use it first.
The GPT-5.6 rollout happened in three distinct stages, and the gap between each one tells its own story about how much friction a government review can add to a product launch that would normally take days, not weeks.
That’s a 13-day gap between the restricted preview and general availability, shorter than the 30-day review window the executive order originally called for. OpenAI has framed the faster timeline as evidence that cooperation works better than confrontation, though the company was careful to note in its own messaging that it sees the arrangement as a one-off rather than a template it wants repeated with every future release.
The Department of Commerce’s Center for AI Standards and Innovation ran the technical evaluation, and OpenAI sent its own engineers to Washington to walk through specific concerns raised mid-review, according to Engadget. That back-and-forth, rather than a single up-or-down decision, is what stretched the process past its original window. It also set a precedent other labs are now measuring themselves against.
GPT-5.6 splits into three tiers, a structure OpenAI has used in prior generations but with new reasoning modes layered on top of the flagship. Sol is built for the hardest agentic and cybersecurity workloads. Terra is the balanced, everyday option. Luna targets high-volume tasks where speed and price matter more than peak capability. A fourth variant, Sol Ultra, adds a multi-subagent coordination mode for teams running complex, long-horizon tasks, though OpenAI has not published standalone pricing for it.
Sol introduces two new reasoning-effort settings on top of the standard mode: a “max” setting for harder single-pass problems, and an “ultra” setting that coordinates multiple subagents on one task. All three tiers ship with improved prompt caching, and cached token reads now get a 90% discount with a 30-minute minimum cache lifetime, a real cost lever for teams running long agentic sessions that repeat context across many calls.
Pricing and tier data compiled from Engadget and the GPT-5.6 entry on Wikipedia, cross-checked against Tech Insider’s own reporting on the June 26 gated launch.
For developers deciding where to point their traffic, the practical choice comes down to three questions: how much reasoning depth does the task need, how sensitive is the workload to latency, and how much does the cache-discount structure change the real cost. Luna at $1 input and $6 output per million tokens is the obvious pick for classification, summarization, and other high-volume jobs where GPT-5.6’s floor is still well above what older, cheaper models could do. Terra, at $2.50 and $15, is OpenAI’s pitch for the bulk of production traffic. The company is explicitly positioning it as roughly half the cost of GPT-5.5 for comparable everyday performance, which puts real pressure on teams still budgeting around last generation’s pricing.
Sol, at $5 and $30, is the tier built for the workloads where mistakes are expensive: agentic coding pipelines, cybersecurity triage, and multi-step research tasks. Sol Ultra sits above that for teams willing to pay for the “ultra” subagent mode on the hardest problems, though the lack of published pricing makes it hard to judge value until OpenAI puts a number on it. The 90% cache discount changes the math meaningfully for any team running repeated, long-context agentic sessions, since a big share of tokens in those workflows are re-read context rather than fresh generation.
Sol’s 88.8% on TerminalBench 2.1 puts it ahead of every publicly available model it’s been directly compared against, and Sol Ultra’s 91.9% extends that lead further. But the gap to the nearest competitor is thin. Anthropic’s Claude Mythos 5 scores 88.0% on the same benchmark, close enough that the ranking could flip with the next point release from either lab. GPT-5.5, last generation’s flagship, scored 83.4%, meaning even Terra, the mid-tier GPT-5.6 model, now beats OpenAI’s own previous top model on this measure.
Benchmark comparisons compiled via the GPT-5.6 Wikipedia entry, which cites TerminalBench 2.1 as the comparison standard across both labs’ recent releases.
TerminalBench evaluates a model’s ability to complete multi-step coding and system administration tasks inside a real terminal environment, rather than answering questions about code in isolation. It rewards models that can plan a sequence of commands, recover from errors, and verify their own output, which is why labs increasingly lead with it when pitching agentic capability rather than older, single-turn benchmarks.
OpenAI didn’t limit its July 9 announcement to text and code models. GPT-Live, a new voice product rolling out as GPT-Live-1 and a smaller GPT-Live-1 mini, launched the same day and is designed to make spoken exchanges with ChatGPT feel closer to a real conversation rather than a question-and-answer loop with noticeable latency between turns. Both versions are rolling out globally to ChatGPT users, and the timing is not an accident. Pairing a flagship reasoning model with a flagship voice product on the same day lets OpenAI frame July 9 as a single, larger platform update rather than just the end of a five-week regulatory delay.
For developers, GPT-Live matters less as a headline feature and more as a signal of where OpenAI expects usage to grow. Voice interfaces carry different latency and cost tradeoffs than text, and bundling a voice launch with GPT-5.6’s tiered pricing structure suggests OpenAI is building toward products where a user might move between text, code, and voice inside a single session without switching models manually.
GPT-5.6 didn’t launch into a vacuum. Anthropic went through a nearly identical government review for its own top-tier models, Mythos and Fable, before receiving permission to redeploy them, per Engadget’s reporting. That’s a meaningful data point on its own: the review process applied to more than one lab, which suggests it’s becoming policy rather than a one-time response to a single company. Anthropic’s Claude Opus 4.8 and Claude Sonnet 5 remain the models most enterprise buyers compare GPT-5.6 against directly, particularly on cost-to-performance for production workloads rather than peak benchmark scores.
Google, meanwhile, has taken a distribution-first approach rather than chasing benchmark leadership. Gemini 3.5 Flash became the new default model inside Google Search starting in May 2026, giving Google a scale advantage that has nothing to do with which lab tops a leaderboard in any given month, according to Google’s own I/O 2026 announcement. Stanford’s 2026 AI Index Report adds useful context to the broader race: as of March 2026, Anthropic’s top model led the best Chinese models by just 2.7 percentage points on benchmark averages, down from a much wider gap a year earlier, with the two sides trading the lead multiple times since early 2025, per the Stanford HAI report.
Open-weight models add a third pressure point. Tech Insider has separately reported on GLM-5.2 undercutting GPT-5.5 on price and on Nvidia’s 550-billion-parameter Nemotron 3 Ultra, both signs that the gap between closed frontier models and open alternatives keeps narrowing on cost even where it hasn’t closed on raw capability.
GPT-5.6 arrives while ChatGPT’s dominance is being chipped at from multiple directions, and the picture looks different depending on which slice of the market gets measured. On raw scale, ChatGPT is still enormous: roughly 1 billion monthly active users as of May 2026, and chatgpt.com pulled in 5.63 billion visits that same month, ranking ninth among all websites in the US, according to Semrush’s traffic data. Measured against all digital queries, including traditional search, ChatGPT holds about 17.9% of the total, per First Page Sage’s Q2 2026 market share report.
Narrow the lens to AI chatbots specifically, and ChatGPT still leads by a wide margin, holding an estimated 68% of AI-chatbot-specific traffic even as Google keeps roughly 90% of traditional search, according to Sedestral’s 2026 market share analysis. The more telling trend shows up in referral data. ChatGPT’s share of B2B AI referral traffic fell from 89% to 63% over an eight-month stretch, while Claude’s share climbed from 1.4% to 18.5% and Gemini’s referral volume roughly quadrupled over the same period, per the 2026 AI Search Traffic Report from Goodie.
None of these figures measure exactly the same thing, and readers should treat them as separate snapshots rather than one consistent trend line. But they point in a similar direction: ChatGPT’s scale advantage in raw traffic remains large, while its share of specific, high-value referral channels is eroding as Claude and Gemini pick up ground.
For enterprise buyers, GPT-5.6’s arrival resets the price-performance conversation just as competitors were closing the gap. Terra’s positioning at half of GPT-5.5’s cost directly targets the mid-market tier where Claude Sonnet 5 and GLM-5.2 have both been making a case on price. That’s a defensive move as much as an offensive one. If OpenAI’s own pricing had stayed flat while rivals undercut it, procurement teams renewing annual contracts this quarter would have had an easy argument to switch.
The regulatory dimension carries its own weight for the industry. A government review process that both OpenAI and Anthropic have now gone through, even briefly, changes how every lab plans a launch calendar. Model release dates that used to be entirely up to the company shipping them now have to account for a federal review window measured in weeks, which pushes labs toward announcing capabilities well ahead of actual availability, exactly the pattern seen with GPT-5.6’s own staged rollout.
There’s a knock-on effect for smaller companies building on top of these models, too. A startup that planned a product launch around GPT-5.6’s original expected release date had to either delay alongside OpenAI or ship on GPT-5.5 and migrate later. That kind of dependency risk was rare before this year, when the biggest planning variable for API-dependent products was usually rate limits or pricing, not a federal review clock sitting between a lab and its own release calendar.
Frontier AI labs have faced plenty of informal pressure before, from voluntary safety commitments to congressional hearings, but GPT-5.6 is the first release to go through a binding, government-managed access list before its public debut. That distinction matters for how the next several launches get planned. Once one administration establishes that it can gate access to a specific model on national security grounds, the precedent doesn’t disappear when that particular review ends.
OpenAI’s own public position, that it doesn’t want this arrangement to become the norm, reads as an attempt to draw a line after the fact rather than prevent the precedent from forming in the first place. Anthropic’s parallel experience with Mythos and Fable suggests the CAISI review isn’t going away as a one-off. Whether it becomes a permanent fixture of frontier model releases or a temporary response to a specific set of 2026 concerns is likely to become clearer with the next major release from either lab.
GPT-5.6 is available in Canada through the same ChatGPT and API channels as the US rollout, with pricing quoted in USD per million tokens rather than a separate Canadian rate. That puts OpenAI’s per-token pricing in the same currency conversation Canadian developers already navigate with AWS, Azure, and other US-billed cloud services. Competing consumer options have taken a different approach: Google AI Plus launched in Canada in January 2026 at roughly CA$10.99 per month, a flat consumer price that sidesteps the token-metering question entirely, as Tech Insider covered in its ranking of AI chatbots for Canadian users.
For Canadian development teams building on the API rather than using the consumer app, the practical decision looks the same as it does anywhere else: Terra’s price cut makes it the likely default for production traffic, while Sol stays reserved for the workloads where the extra cost is easy to justify against the alternative of a failed agentic task.
What is GPT-5.6, and when did it launch?
GPT-5.6 is OpenAI’s latest model family, made up of three tiers: Sol, Terra, and Luna. It shipped in restricted form to about 20 vetted organizations on June 26, 2026, then became generally available to all ChatGPT, API, and Codex users on July 9, 2026.
Why was GPT-5.6’s public release delayed?
An AI cybersecurity executive order signed in early June 2026 required frontier labs to submit their most capable models for a federal security review before public release. The Department of Commerce’s Center for AI Standards and Innovation conducted that review for GPT-5.6.
What are Sol, Terra, and Luna?
They’re the three GPT-5.6 tiers. Sol is the flagship, built for agentic coding and cybersecurity work. Terra is the balanced, everyday option priced at roughly half of GPT-5.5. Luna is the fastest, lowest-cost tier for high-volume tasks. A fourth variant, Sol Ultra, adds a multi-subagent reasoning mode.
How much does GPT-5.6 cost?
Sol costs $5 per million input tokens and $30 per million output tokens. Terra costs $2.50 input and $15 output. Luna costs $1 input and $6 output. Sol Ultra’s pricing has not been published.
Is GPT-5.6 available in Canada?
Yes. Canadian users access GPT-5.6 through the same ChatGPT and API channels as US users, with per-token pricing billed in USD.
How does GPT-5.6 compare to Claude and Gemini?
On TerminalBench 2.1, GPT-5.6 Sol scores 88.8%, narrowly ahead of Anthropic’s Claude Mythos 5 at 88.0%. Google has focused less on topping benchmarks and more on distribution, making Gemini 3.5 Flash the default model inside Google Search since May 2026.
What happened to the 20 organizations that had early access?
They keep their existing access under the same terms. General availability adds every other ChatGPT, Codex, and API user on top of that early group rather than replacing it.
Will other AI labs face the same government review?
Anthropic already went through a comparable review for its Mythos and Fable models before receiving clearance to redeploy them, which suggests the process is being applied across labs rather than singling out OpenAI.
Marcus Chen is a Senior Tech Reporter at Tech Insider covering cloud computing, enterprise software, and the business of technology. Before joining TI, he spent five years at ZDNet covering digital transformation across European enterprises and three years at The Register reporting on cloud infrastructure. Marcus is known for his deep dives into cloud cost optimization and multi-cloud strategy. He holds a degree in Computer Science from Imperial College London and speaks regularly at KubeCon and CloudNative events.
Tech Insider delivers in-depth coverage of the technologies shaping the future: AI, cybersecurity, cloud computing, hardware, and the trends that matter.
