Creators have compared Claude Opus and GPT-5 directly in 4 videos. Claude Opus leans positive across 45 videos; GPT-5 is more positive across 5 videos.
| Date | Channel | Video |
|---|---|---|
| 12 Jul 2026 | WorldofAI | Claude Opus 5 LEAKS, GPT-6 ALREADY, Kimi K3 Soon, Fable 5.1, NEO Hands, & More! AI NEWS |
| 11 Jul 2026 | Jack Roberts | 5 Insane ChatGPT 5.6 Sol Use Cases... |
| 6 Jul 2026 | Jack Roberts | Fable 5 Agentic OS is Insane... just watch |
| 15 Jun 2026 | Build Great Products | Fable 5 Might Never Come Back. Here's What to Do Next |
Get every new Claude Opus and GPT-5 video summarised in your inbox.
Try it freeSeveral creators who directly compared Claude Opus and GPT-5 found GPT-5.6 Sol outperforming Claude Opus 4.8 on formal coding benchmarks. Wes Roth reports that GPT-5.6 Soul scored 53.6% on Agents Last Exam versus Claude Fable 5 at 40.5%, and achieved a state-of-the-art score of 80 on the Artificial Analysis coding agent index, beating Fable 5 by 2.8 points. Matthew Berman similarly notes that GPT-5.6 Soul takes the most direct path to a solution of any model he has used, consuming far fewer tokens per task than Claude Fable, making its real cost-per-task significantly lower despite similar per-token list pricing.
Creators are more measured when discussing Claude Opus on its own terms, however. WorldofAI notes that Claude Opus 4 scores 80.4% on SWE-Bench Pro, a figure that Grok 4.5 and GPT-5.5 struggle to match in that specific benchmark, suggesting Opus retains an edge on certain hard coding tasks. Jack Roberts observed in a direct side-by-side test that Claude Fable 5 produced slightly better output than Claude Opus 4.8 on cold email and website design, but the difference was subtle enough that most users would not notice — reinforcing the view among several creators that the headline capability gap between the top Claude and GPT-5 models is narrower than marketing suggests.
The picture is further complicated by reliability concerns around Claude Opus. Riley Brown and WorldofAI both report that the re-released Claude Opus 5 carries a safety classifier that automatically reroutes flagged requests to Claude Opus 4.8, meaning benchmark scores and real-world sessions may be measuring a mix of two different models. WorldofAI documents a roughly ten-point drop in Anthropic's own Apex Sway benchmark scores after redeployment, while GPT-5.6 attracted no equivalent reliability caveats from the same creators.
Matthew Berman's extended testing of GPT-5.6 in Codex is among the most striking evidence in the corpus for GPT-5's agentic depth. He describes running unattended Codex loops for up to six days to produce a Minecraft clone with biomes and NPCs, a fully functional Excel clone, and a physics game — with GPT-5.6 autonomously directing browser control to click around the real Excel application during the build. Berman also used GPT-5.6-powered Codex to automate DNS migrations across multiple hosting providers, framing it as genuinely autonomous infrastructure management rather than assisted scripting.
Creators who discuss Claude Opus in agentic contexts tend to position it as an orchestrator rather than a sole autonomous executor. Jack Roberts demonstrates Claude Opus 4 sitting at the top of a Ministry of Agents stack, directing sub-agents including GPT-5.5, DeepSeek, and GLM 5.2, with the key insight being that Opus provides the taste and strategic direction while cheaper models handle volume work. A separate video from Creator Magic shows Claude Opus 4.8 used to figure out a complex task once — such as bypassing Reddit bot-blocking — saving the result as a reusable skill, then handing repetitive execution to Claude Haiku 4.5, a pattern that treats Opus as an expensive but valuable problem-solver rather than an always-on worker.
Dan Shipper, who tested GPT-5.6 in Codex for thirty days, describes it as his primary operating system for all knowledge work and explicitly calls it more practical than Claude Opus 4 for everyday use. He built a personal email triage app, a company pulse feed, and a SaaS maintenance badge tool entirely within this GPT-5.6 stack. The contrast with Claude Opus is notable: creators tend to deploy Opus for high-stakes architectural decisions and overnight autonomous review tasks, whereas GPT-5.6 in Codex is described as a continuous daily driver for a wider range of unattended workflows.
Pricing is one of the sharpest points of contrast in the corpus, and creators who discuss both models directly are consistent in their view that GPT-5.6 Sol delivers superior value per task. Wes Roth cites the Agents Last Exam result as a concrete illustration: GPT-5.6 Sol achieved a better score than Claude Fable 5 at one-third the cost — $763 versus $2,300. Matthew Berman adds that GPT-5.6 is priced identically to GPT-5.5 at $5 per million input tokens and $30 per million output tokens, yet uses far fewer output tokens per task than Claude Fable, meaning the practical per-task cost is substantially lower despite matching list prices.
On the Claude Opus side, creators highlight extreme session costs as a real friction point for builders. Riley Brown reports a single four-prompt coding session costing $174 with Claude Opus 5, and a single analysis prompt costing $135. He also documents a session where $242 of a $321 charge went to Claude Opus 4.8 due to the safety-classifier rerouting behaviour, making cost predictability a genuine concern. Jack Roberts frames Claude Fable 5 as costing roughly $45 per 2 million tokens in and 500,000 tokens out, and recommends reserving it strictly for taste-sensitive or strategically irreversible decisions while routing everything else to cheaper models.
Several creators converge on a model-routing strategy that implicitly treats Claude Opus and GPT-5 as expensive planning layers rather than default execution models. Matthew Berman calculates that using a frontier model like Claude Opus for everything costs approximately $9.50 per feature, versus roughly $3.20 when execution is offloaded to a cheaper model — a 68% saving. Jack Roberts demonstrates that Claude Opus 4.8 paired with a multi-model swarm produces competitive results against solo Claude Fable 5 at a fraction of the price, and a parallel pattern emerges for GPT-5.6 Sol, which Jack Roberts separately describes as delivering near-frontier intelligence at one-third the cost of GPT-5.
GPT-5.6's primary integration story in the corpus centres on Codex and the newly launched ChatGPT Work desktop application, which OpenAI positions as analogous to Claude Code for non-developers as well as engineers. Wes Roth notes that GPT-5.6 Soul autonomously post-trained the smaller Luna model using a short Codex prompt — a recursive self-improvement milestone that several creators treat as evidence of GPT-5's deep integration with its own development toolchain. Matthew Berman highlights Codex browser control as a standout feature, using it to automate DNS migrations and auto-scale a Supabase instance, tasks that required no manual intervention.
Claude Opus's integration story is more distributed across third-party tools. Creators describe Claude Code built on Claude Opus 4.8 as the primary coding agent across a range of builds: Creator Magic uses it to scaffold and validate projects live, Jack Roberts builds Clay MCP-powered sales prospecting workflows inside Claude Code, and one creator uses it to restructure a unified memory system into a searchable knowledge explorer. WorldofAI briefly spotted Claude Honeycomb — likely Claude Opus 5 — inside Cursor with a one-million-token context window and an extra-high reasoning mode, though the same creator notes early results looked underwhelming compared to current state-of-the-art models.
A direct comparison emerges in the corpus between Codex and Claude Code as competing agentic IDE environments. Dan Shipper describes Codex with GPT-5.6 as his full operating system for knowledge work and calls it more practical than Claude Opus for everyday use, citing its continuous context via the Codex Chronicle screenshot feed. By contrast, creators building on Claude Code emphasise its strength as an orchestration layer over other models and tools rather than as a standalone executor, suggesting the two environments suit different builder archetypes — those wanting a self-contained long-running agent versus those wanting a flexible multi-model conductor.
Access stability is one of the starkest differentiators the corpus reveals. Claude Opus 5 was suspended 78 hours after its June 2026 launch due to a US government national security directive, then redeployed with stricter safety classifiers and nationality verification requirements, according to Riley Brown. WorldofAI documents the practical consequence: a safety classifier that automatically reroutes flagged requests to Claude Opus 4.8 mid-session, making it difficult for builders to know which model is actually handling their work at any given moment. Plan-included access then ended on 7 July 2026, switching to credit billing at costs that Riley Brown describes as extreme.
Creators report no equivalent access or rerouting issues with GPT-5.6 Sol during the same period. Matthew Berman ran six-day unattended Codex loops without interruption, and Dan Shipper describes thirty days of continuous use as his primary operating system. The Build Great Products channel frames the contrast explicitly, noting that even with Claude Opus 5 offline, GPT-5 remained available and powerful enough to build businesses significantly faster than pre-AI workflows — treating GPT-5 as the stable fallback when Opus was unavailable.
Several creators flag that the Claude Opus reliability story is not purely negative: Creator Magic notes Anthropic temporarily restored Claude Sonnet 5 to Max and Pro plans as a goodwill gesture during the disruption, and Jack Roberts continues to recommend Claude Opus 4.8 as the orchestrating model in multi-agent pipelines for its taste and strategic judgement. Nevertheless, the corpus reflects a period in which builders working on time-sensitive agentic projects were actively seeking alternatives — with GPT-5.6 Sol and Grok 4.5 both cited as beneficiaries of that search.
Creators are divided. Matthew Berman and Dan Shipper describe GPT-5.6 Sol in Codex as the more practical daily driver for long-running agentic tasks, citing its ability to run unattended for days and its lower token consumption per task. Jack Roberts, however, positions Claude Opus 4.8 as the preferred orchestrating model in multi-agent pipelines, arguing its strategic judgement and 'taste' make it valuable for directing cheaper sub-agents — a different kind of agentic role that creators say GPT-5.6 does not obviously replicate.
Multiple creators who tested both models directly conclude that GPT-5.6 Sol is substantially cheaper per task in practice. Wes Roth cites a benchmark result showing GPT-5.6 Sol outperforming Claude Fable 5 at one-third the cost, and Matthew Berman notes GPT-5.6 uses far fewer output tokens per task than Claude Fable despite comparable list pricing. Riley Brown reports Claude Opus 5 API sessions costing $135–$174 for single tasks, and flags that safety-classifier rerouting makes costs unpredictable. Creators consistently recommend routing rather than using either model for all tasks.
Several creators favour Claude Opus as the orchestrating layer. Jack Roberts demonstrates Claude Opus 4 directing a Ministry of Agents stack over GPT-5.5, DeepSeek, and GLM sub-agents, arguing that Opus provides the quality threshold needed to keep multi-agent consensus meaningful. A separate Jack Roberts video reinforces this, finding that a swarm of cheap models without an intelligent orchestrator performs poorly, and that Opus 4.8 fills that orchestrating role effectively. Creators who use GPT-5.6 tend to run it as a self-contained executor in Codex rather than as an orchestrator over other models.
The corpus contains limited direct comparison on this dimension. WorldofAI reports that Claude Honeycomb — likely Claude Opus 5 — was briefly spotted in Cursor with a one-million-token context window and an extra-high reasoning mode, though the same creator notes early results looked underwhelming. GPT-5.6 creators describe Codex Chronicle — a local screenshot feed — as a practical continuous-context mechanism for GPT-5.6, but no specific context window figure for GPT-5.6 is cited in the corpus. Creators building on Claude Code note its ability to share context across tools via a unified memory layer, but this is a tooling feature rather than a native model capability.
Creators broadly treat Claude Opus and GPT-5 as equivalent starting points for beginners. Wes Roth recommends getting a paid Claude or ChatGPT account and maximising daily usage to build and ship small apps, without favouring one over the other. The Build Great Products channel states explicitly that both Claude Opus 4 and GPT-5 are powerful enough to build businesses significantly faster than pre-AI workflows, even when more advanced models are unavailable. The practical recommendation from several creators is to pick one, learn it deeply, and focus on shipping rather than optimising model choice at the outset.
Following Claude Opus and GPT-5 news across YouTube?
summree watches the channels covering Claude Opus and GPT-5 and emails you a summary every time a new video drops. Add your channels once — never miss a release again.
Try it free